Photanima

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3272949" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3272949
 </a>



<h2>
 About this version
</h2>

 Replaced distillation stack with the new
 <a href="https://civitai.red/models/2458426/anima" rel="ugc" target="_blank">
 turbo-v1.1
 </a>
 .


 Use ER-SDE sampler. The
 <code>
 ODE
 </code>
 solver type can improve detail, but it also tends to make people look uglier in this version. Try with and without.


 First impressions: v2.4 is sharper and more detailed, but it is less cinematic and arguably harder to prompt. Also...
 <code>
 (crossed eyes, ugly face:-1)
 </code>
 feels like it's more important than it ought to be. &gt;_&gt;

v2.4 Turbo

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3183251" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3183251
 </a>




 About this version


 Intended for further finetuning or LoRAs. As usual, Turbo is better for inference.




 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>



<hr/>


<h2>
 
 🤓 Technical details
 
</h2>



 v2 is trained on ~2000 images for 45,000 steps. This is an expansion of my
 <a href="https://civitai.red/models/2068592/snakebite-2?modelVersionId=2426502" rel="ugc" target="_blank">
 Snakebite 2.3
 </a>
 dataset with around 700 new images and captions reworked for Anima. Training took approximately 48 hours on a Geforce 3090.




 
 Pros:
 

<ul>
 <li>
 
 Extremely fast.
 
 </li>
 <li>
 
 Extremely good prompt adherence.
 
 </li>
 <li>
 
 Anatomy is pretty stable. If it screws something up, changing your steps by +1/-1 usually fixes it.
 
 </li>
 <li>
 
 Supports up to nearly 2MP with little-to-no distortions.
 
 </li>
</ul>



 At first, I noticed that Photanima's style was inconsistent - it had a tendency to regress toward a cartoony/CGI look as my prompts became more complex. I was able to mostly overcome this by splitting Photanima into constituent content, style-early, and style-late blocks, then boosted the style blocks well past a strength of 1.


 "Style-late" maps to blocks 7, 8, and 9 - these do alter composition to a degree, so we can't boost them as hard as "style-early."


 Images are pretty consistent now, but there are some notable drawbacks.




 
 Cons in v2:
 

<ul>
 <li>
 
 It loses a little knowledge of certain artistic terms like
 <code>
 silhouette
 </code>
 .
 
 </li>
 <li>
 
 Microdetail quality is somewhere between SDXL and ZIT. Honestly, it's really good for a 2B model. Two-step upscaling with Anima doesn't help much, but I'm sure the results would be amazing if you sent a Photanima image to a different model for refinement. Or if that's too much work: just add a little film grain. It does wonders and requires no extra VRAM.
 
 </li>
 <li>
 
 Text capabilities are not as good as those of base Anima. Anything beyond 3 or 4 words is likely going to require numerous re-rolls. This is at least partly due to the Turbo LoRA.
 
 </li>
 <li>
 
 Excessive fluff tags like
 <code>
 masterpiece, absurdres, hyperreal
 </code>
 tend to fry the image. The model is photographic and highly aesthetic by default, so there's no need to drive it harder in that direction.
 
 </li>
</ul>




<hr/>


<h2>
 
 🛠️ Recommended Settings
 
</h2>



 
 v2.3 Turbo:
 

<ul>
 <li>
 
 8-9 steps.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 CFG exactly 1.
 
 </li>
 <li>
 
 Preferred resolution: 1040x1520 or 832x1216.
 
 </li>
 <li>
 
 For maximum realism, begin your prompt with
 <code>
 real life photo
 </code>
 . If that's not enough, add
 <code>
 photo \(medium\)
 </code>
 and increase its strength until satisfied. You can usually go up to a crazy strength value like 5 or 6 without breaking the image.
 
 </li>
 <li>
 
 You can remove NegPip fluff to improve anatomy (e.g. fingers) at the cost of some photographic texture.
 
 </li>
 <li>
 
 Workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/CK806jd6/image-photanima-turbo-v2-3.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Workflow with zero custom nodes (sacrifices realism and speed):
 <a href="https://i.ibb.co/tpMzvxJY/image-photanima-turbo-v2-3-no-custom-nodes.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 v2.2 Turbo:
 

<ul>
 <li>
 
 6-8 steps. Images often look best at 6, but anatomy is more stable at 8-10, especially with complex prompts.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 You can reduce the first number on the sigma curve to 0.95-0.99 to improve realism. This reduces saturation and adds a little noise, but makes the model less stable.
 
 </li>
 <li>
 
 Workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/JjhHv266/image-photanima-turbo-v2-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Simple workflow with fewer custom nodes:
 <a href="https://i.ibb.co/p6t1YcL3/image-photanima-turbo-v2-0-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 v2.2 Base/Non-Turbo:
 

<ul>
 <li>
 
 You can get a good image in 25 steps, but 40 is often better.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG between 3.5 to 4.
 
 </li>
 <li>
 
 Recommended fluff: "(photo \(medium\):1), real life, score_9, aesthetic"
 
 </li>
 <li>
 
 Recommended negative prompt: "toon \(style\), anime coloring, painting \(medium\), airbrushed, mutation, distortion, ai-assisted, glossy, shiny, shiny skin, worst quality, score_3, score_4"
 
 </li>
 <li>
 
 I have found it's helpful to decay conditioning strength from 2 to 1 over the first ~40% of steps. The stock workflow does this.
 
 </li>
 <li>
 
 Workflow optimized for realism:
 <a href="https://i.ibb.co/xtrs0F5Q/image-photanima-v2-1b.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>


<hr/>


<h2>
 
 🗺️ Roadmap
 
</h2>



 I'm pretty excited about the potential of Anima, but let's be clear: I'm not claiming that this checkpoint is a "ZIT killer." The correct model to compare this against is SDXL/IL - and I'm confident that Anima can dethrone it with enough community effort.


 Directions I'd like to explore next:

<ul>
 <li>
 
 (✅ Done in v2) There are a handful of Anima "detailer" LoRAs on Civitai. These are not intended for photography, but with enough block pruning, you never know. The right mix could go a long way.
 
 </li>
 <li>
 
 I suspect further increasing the dataset to ~3k images would help resolve remaining issues related to certain textures or model biases.
 
 </li>
 <li>
 
 (✅ Done in v2) I'm eagerly awaiting the release of Anima Turbo 1.0. The current Turbo solution is based on Preview3 and I think it's holding back this model's potential a little.
 
 </li>
 <li>
 
 I'm also looking forward to Anima support in OneTrainer. It will make trying experimental configs a lot less of a hassle compared to kohya-ss. For this v1 run, I stuck with safe values (prodigy, 1.0 LR, no fancy flags.)
 
 </li>
</ul>

v2.3

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3112450" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3112450
 </a>



<h2>
 About this version
</h2>

 
 Model changes:
 

<ul>
 <li>
 
 Replaced Turbo LoRA v0.2 with the new
 <a href="https://civitai.red/models/2458426/anima?modelVersionId=3108589" rel="ugc" target="_blank">
 turbo-v1.0 checkpoint
 </a>
 . This has significantly improved textures but slightly weaker aesthetics.
 
 </li>
 <li>
 
 A small amount of the previous Turbo LoRA was merged back in to help stabilize the image.
 
 </li>
 <li>
 
 Refined my approach to block pruning for smoother mathematical curve. In theory, this should reduce distortions caused by sharp changes across block strengths.
 
 </li>
 <li>
 
 Added a splash of Blocks 11 and 12 from my custom training as they help tame colors without radically affecting composition.
 
 </li>
</ul>

 
 Workflow changes:
 

<ul>
 <li>
 
 Switched from
 <code>
 qwen_image_vae
 </code>
 to
 <code>
 wan_2.1_vae
 </code>
 for slightly better microdetail.
 
 </li>
 <li>
 
 INT8 conversion path now uses the
 <code>
 convrot
 </code>
 outlier method for improved quality.
 
 </li>
 <li>
 
 New custom sigma curve. It is just
 <code>
 beta
 </code>
 scheduler with the first value bumped slightly above 1.0. This is an easy way to add detail.
 
 </li>
 <li>
 
 Slight adjustments to default fluff prompt.
 
 </li>
</ul>

 
 Other notes:
 

<ul>
 <li>
 
 Previous versions of Photanima converged at 6 steps, but v2.3 needs at least 8 steps.
 
 </li>
</ul>



 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>

v2.3 Turbo

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3098914" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3098914
 </a>




 About this version


 Non-Turbo version for training. As always, I recommend Turbo for inference.




 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>



<hr/>


<h2>
 
 🤓 Technical details
 
</h2>



 v2 is trained on ~2000 images for 45,000 steps. This is an expansion of my
 <a href="https://civitai.red/models/2068592/snakebite-2?modelVersionId=2426502" rel="ugc" target="_blank">
 Snakebite 2.3
 </a>
 dataset with around 700 new images and captions reworked for Anima. Training took approximately 48 hours on a Geforce 3090.




 
 Pros:
 

<ul>
 <li>
 
 Extremely fast.
 
 </li>
 <li>
 
 Extremely good prompt adherence.
 
 </li>
 <li>
 
 Anatomy is pretty stable. If it screws something up, changing your steps by +1/-1 usually fixes it.
 
 </li>
 <li>
 
 Supports up to nearly 2MP with little-to-no distortions.
 
 </li>
</ul>



 At first, I noticed that Photanima's style was inconsistent - it had a tendency to regress toward a cartoony/CGI look as my prompts became more complex. I was able to mostly overcome this by splitting Photanima into constituent content, style-early, and style-late blocks, then boosted the style blocks well past a strength of 1.


 "Style-late" maps to blocks 7, 8, and 9 - these do alter composition to a degree, so we can't boost them as hard as "style-early."


 Images are pretty consistent now, but there are some notable drawbacks.




 
 Cons in v2:
 

<ul>
 <li>
 
 It loses a little knowledge of certain artistic terms like
 <code>
 silhouette
 </code>
 .
 
 </li>
 <li>
 
 Microdetail quality is somewhere between SDXL and ZIT. Honestly, it's really good for a 2B model. Two-step upscaling with Anima doesn't help much, but I'm sure the results would be amazing if you sent a Photanima image to a different model for refinement. Or if that's too much work: just add a little film grain. It does wonders and requires no extra VRAM.
 
 </li>
 <li>
 
 Text capabilities are not as good as those of base Anima. Anything beyond 3 or 4 words is likely going to require numerous re-rolls. This is at least partly due to the Turbo LoRA.
 
 </li>
 <li>
 
 Excessive fluff tags like
 <code>
 masterpiece, absurdres, hyperreal
 </code>
 tend to fry the image. The model is photographic and highly aesthetic by default, so there's no need to drive it harder in that direction.
 
 </li>
</ul>




<hr/>


<h2>
 
 🛠️ Recommended Settings (for latest versions)
 
</h2>



 
 Turbo:
 

<ul>
 <li>
 
 6-8 steps. Images often look best at 6, but anatomy is more stable at 8-10, especially with complex prompts.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG exactly 1.
 
 </li>
 <li>
 
 Preferred resolution: 1040x1520 or 832x1216.
 
 </li>
 <li>
 
 For maximum realism, begin your prompt with
 <code>
 real life photo
 </code>
 . If that's not enough, add
 <code>
 photo \(medium\)
 </code>
 and increase its strength until satisfied. You can usually go up to a crazy strength value like 5 or 6 without breaking the image.
 
 </li>
 <li>
 
 You can reduce the first number on the sigma curve to 0.95-0.99 to improve realism. This reduces saturation and adds a little noise, but makes the model less stable.
 
 </li>
 <li>
 
 You can remove NegPip fluff to improve anatomy (e.g. fingers) at the cost of some photographic texture.
 
 </li>
 <li>
 
 Newest workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/JjhHv266/image-photanima-turbo-v2-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Simple workflow with fewer custom nodes:
 <a href="https://i.ibb.co/p6t1YcL3/image-photanima-turbo-v2-0-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 Base/Non-Turbo:
 

<ul>
 <li>
 
 You can get a good image in 25 steps, but 40 is often better.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG between 3.5 to 4.
 
 </li>
 <li>
 
 Recommended fluff: "(photo \(medium\):1), real life, score_9, aesthetic"
 
 </li>
 <li>
 
 Recommended negative prompt: "toon \(style\), anime coloring, painting \(medium\), airbrushed, mutation, distortion, ai-assisted, glossy, shiny, shiny skin, worst quality, score_3, score_4"
 
 </li>
 <li>
 
 I have found it's helpful to decay conditioning strength from 2 to 1 over the first ~40% of steps. The stock workflow does this.
 
 </li>
 <li>
 
 Newest workflow optimized for realism:
 <a href="https://i.ibb.co/xtrs0F5Q/image-photanima-v2-1b.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>


<hr/>


<h2>
 
 🗺️ Roadmap
 
</h2>



 I'm pretty excited about the potential of Anima, but let's be clear: I'm not claiming that this checkpoint is a "ZIT killer." The correct model to compare this against is SDXL/IL - and I'm confident that Anima can dethrone it with enough community effort.


 Directions I'd like to explore next:

<ul>
 <li>
 
 (✅ Done in v2) There are a handful of Anima "detailer" LoRAs on Civitai. These are not intended for photography, but with enough block pruning, you never know. The right mix could go a long way.
 
 </li>
 <li>
 
 I suspect further increasing the dataset to ~3k images would help resolve remaining issues related to certain textures or model biases.
 
 </li>
 <li>
 
 (✅ Done in v2) I'm eagerly awaiting the release of Anima Turbo 1.0. The current Turbo solution is based on Preview3 and I think it's holding back this model's potential a little.
 
 </li>
 <li>
 
 I'm also looking forward to Anima support in OneTrainer. It will make trying experimental configs a lot less of a hassle compared to kohya-ss. For this v1 run, I stuck with safe values (prodigy, 1.0 LR, no fancy flags.)
 
 </li>
</ul>


<hr/>



 Thank you. As always, I look forward to your feedback. Please share the model and upload some images to help it gain traction.

v2.2

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3064104" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3064104
 </a>



<h2>
 About this version
</h2>

 Improves photographic texture and is more likely to remain stable at 9-12 steps (although I still prefer 6 steps in many cases!)


 This is an experimental update. I collected a new dataset of ~200 images with the goal of addressing weaknesses in the previous version. It was mainly comprised of food, animals, and various texture close-ups. I wanted to improve Photanima's material rendering across a range of lighting conditions.


 I trained these for 14k steps on top of v2.1 as a "stimulus pack," instead of retraining the model from scratch. I then pruned the resulting LoRA, keeping only the style blocks.


 The approach shows promise, but it remains to be seen whether it will work for more challenging problems, such as anatomy or text.




 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>



<hr/>


<h2>
 
 🤓 Technical details
 
</h2>



 v2 is trained on ~2000 images for 45,000 steps. This is an expansion of my
 <a href="https://civitai.red/models/2068592/snakebite-2?modelVersionId=2426502" rel="ugc" target="_blank">
 Snakebite 2.3
 </a>
 dataset with around 700 new images and captions reworked for Anima. Training took approximately 48 hours on a Geforce 3090.




 
 Pros:
 

<ul>
 <li>
 
 Extremely fast.
 
 </li>
 <li>
 
 Extremely good prompt adherence.
 
 </li>
 <li>
 
 Anatomy is pretty stable. If it screws something up, changing your steps by +1/-1 usually fixes it.
 
 </li>
 <li>
 
 Supports up to nearly 2MP with little-to-no distortions.
 
 </li>
</ul>



 At first, I noticed that Photanima's style was inconsistent - it had a tendency to regress toward a cartoony/CGI look as my prompts became more complex. I was able to mostly overcome this by splitting Photanima into constituent content, style-early, and style-late blocks, then boosted the style blocks well past a strength of 1.


 "Style-late" maps to blocks 7, 8, and 9 - these do alter composition to a degree, so we can't boost them as hard as "style-early."


 Images are pretty consistent now, but there are some notable drawbacks.




 
 Cons in v2:
 

<ul>
 <li>
 
 It loses a little knowledge of certain artistic terms like
 <code>
 silhouette
 </code>
 .
 
 </li>
 <li>
 
 Microdetail quality is somewhere between SDXL and ZIT. Honestly, it's really good for a 2B model. Two-step upscaling with Anima doesn't help much, but I'm sure the results would be amazing if you sent a Photanima image to a different model for refinement. Or if that's too much work: just add a little film grain. It does wonders and requires no extra VRAM.
 
 </li>
 <li>
 
 Text capabilities are not as good as those of base Anima. Anything beyond 3 or 4 words is likely going to require numerous re-rolls. This is at least partly due to the Turbo LoRA.
 
 </li>
 <li>
 
 Excessive fluff tags like
 <code>
 masterpiece, absurdres, hyperreal
 </code>
 tend to fry the image. The model is photographic and highly aesthetic by default, so there's no need to drive it harder in that direction.
 
 </li>
</ul>




<hr/>


<h2>
 
 🛠️ Recommended Settings (for latest versions)
 
</h2>



 
 Turbo:
 

<ul>
 <li>
 
 6-8 steps. Images often look best at 6, but anatomy is more stable at 8-10, especially with complex prompts.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG exactly 1.
 
 </li>
 <li>
 
 Preferred resolution: 1040x1520 or 832x1216.
 
 </li>
 <li>
 
 For maximum realism, begin your prompt with
 <code>
 real life photo
 </code>
 . If that's not enough, add
 <code>
 photo \(medium\)
 </code>
 and increase its strength until satisfied. You can usually go up to a crazy strength value like 5 or 6 without breaking the image.
 
 </li>
 <li>
 
 You can reduce the first number on the sigma curve to 0.95-0.99 to improve realism. This reduces saturation and adds a little noise, but makes the model less stable.
 
 </li>
 <li>
 
 You can remove NegPip fluff to improve anatomy (e.g. fingers) at the cost of some photographic texture.
 
 </li>
 <li>
 
 Newest workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/JjhHv266/image-photanima-turbo-v2-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Simple workflow with fewer custom nodes:
 <a href="https://i.ibb.co/p6t1YcL3/image-photanima-turbo-v2-0-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 Base/Non-Turbo:
 

<ul>
 <li>
 
 You can get a good image in 25 steps, but 40 is often better.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG between 3.5 to 4.
 
 </li>
 <li>
 
 Recommended fluff: "(photo \(medium\):1), real life, score_9, aesthetic"
 
 </li>
 <li>
 
 Recommended negative prompt: "toon \(style\), anime coloring, painting \(medium\), airbrushed, mutation, distortion, ai-assisted, glossy, shiny, shiny skin, worst quality, score_3, score_4"
 
 </li>
 <li>
 
 I have found it's helpful to decay conditioning strength from 2 to 1 over the first ~40% of steps. The stock workflow does this.
 
 </li>
 <li>
 
 Newest workflow optimized for realism:
 <a href="https://i.ibb.co/xtrs0F5Q/image-photanima-v2-1b.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>


<hr/>


<h2>
 
 🗺️ Roadmap
 
</h2>



 I'm pretty excited about the potential of Anima, but let's be clear: I'm not claiming that this checkpoint is a "ZIT killer." The correct model to compare this against is SDXL/IL - and I'm confident that Anima can dethrone it with enough community effort.


 Directions I'd like to explore next:

<ul>
 <li>
 
 (✅ Done in v2) There are a handful of Anima "detailer" LoRAs on Civitai. These are not intended for photography, but with enough block pruning, you never know. The right mix could go a long way.
 
 </li>
 <li>
 
 I suspect further increasing the dataset to ~3k images would help resolve remaining issues related to certain textures or model biases.
 
 </li>
 <li>
 
 (✅ Done in v2) I'm eagerly awaiting the release of Anima Turbo 1.0. The current Turbo solution is based on Preview3 and I think it's holding back this model's potential a little.
 
 </li>
 <li>
 
 I'm also looking forward to Anima support in OneTrainer. It will make trying experimental configs a lot less of a hassle compared to kohya-ss. For this v1 run, I stuck with safe values (prodigy, 1.0 LR, no fancy flags.)
 
 </li>
</ul>


<hr/>



 Thank you. As always, I look forward to your feedback. Please share the model and upload some images to help it gain traction.

v2.2 Turbo

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3013998" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3013998
 </a>



<h2>
 About this version
</h2>

 Improves photographic texture by incorporating style blocks from the RealCosplay checkpoint. It's a neat model that aims to enhance both realism and illustrative generations - check it out
 <a href="https://civitai.red/models/2671085/realcosplayandanimeand3d-all-in-one-animabase" rel="ugc" target="_blank">
 here
 </a>
 .


 I also made meaningful improvements to the stock workflow:

<ul>
 <li>
 
 Optimized custom sigma curve.
 
 </li>
 <li>
 
 Introduces NegPip with a "pseudo negative prompt" that works at 1 CFG.
 
 </li>
 <li>
 
 Uses the
 <a href="https://github.com/SparknightLLC/ComfyUI-ConditioningMultiplyAdvanced" rel="ugc" target="_blank">
 Conditioning Multiply Advanced node
 </a>
 to drop conditioning strength to 0 after a certain number of steps, improving texture.
 
 </li>
</ul>

 As a result: v2.1 is the most realistic version of Photanima to date, but anatomy is a little less stable than v2.0. I view it as a worthwhile tradeoff.




 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>



<hr/>


<h2>
 
 🤓 Technical details
 
</h2>



 v2 is trained on ~2000 images for 45,000 steps. This is an expansion of my
 <a href="https://civitai.red/models/2068592/snakebite-2?modelVersionId=2426502" rel="ugc" target="_blank">
 Snakebite 2.3
 </a>
 dataset with around 700 new images and captions reworked for Anima. Training took approximately 48 hours on a Geforce 3090.




 
 Pros:
 

<ul>
 <li>
 
 Extremely fast.
 
 </li>
 <li>
 
 Extremely good prompt adherence.
 
 </li>
 <li>
 
 Anatomy is pretty stable. If it screws something up, changing your steps by +1/-1 usually fixes it.
 
 </li>
 <li>
 
 Supports up to nearly 2MP with little-to-no distortions.
 
 </li>
</ul>



 At first, I noticed that Photanima's style was inconsistent - it had a tendency to regress toward a cartoony/CGI look as my prompts became more complex. I was able to mostly overcome this by splitting Photanima into constituent content, style-early, and style-late blocks, then boosted the style blocks well past a strength of 1.


 "Style-late" maps to blocks 7, 8, and 9 - these do alter composition to a degree, so we can't boost them as hard as "style-early."


 Images are pretty consistent now, but there are some notable drawbacks.




 
 Cons in v2:
 

<ul>
 <li>
 
 It loses a little knowledge of certain artistic terms like
 <code>
 silhouette
 </code>
 .
 
 </li>
 <li>
 
 Microdetail quality is somewhere between SDXL and ZIT. Honestly, it's really good for a 2B model. Two-step upscaling with Anima doesn't help much, but I'm sure the results would be amazing if you sent a Photanima image to a different model for refinement. Or if that's too much work: just add a little film grain. It does wonders and requires no extra VRAM.
 
 </li>
 <li>
 
 Text capabilities are not as good as those of base Anima. Anything beyond 3 or 4 words is likely going to require numerous re-rolls. This is at least partly due to the Turbo LoRA.
 
 </li>
 <li>
 
 Excessive fluff tags like
 <code>
 masterpiece, absurdres, hyperreal
 </code>
 tend to fry the image. The model is photographic and highly aesthetic by default, so there's no need to drive it harder in that direction.
 
 </li>
</ul>




<hr/>


<h2>
 
 🛠️ Recommended Settings (for latest versions)
 
</h2>



 
 Turbo:
 

<ul>
 <li>
 
 6-8 steps. Images often look best at 6, but anatomy is more stable at 8-10, especially with complex prompts.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG exactly 1.
 
 </li>
 <li>
 
 Preferred resolution: 1040x1520 or 832x1216.
 
 </li>
 <li>
 
 For maximum realism, begin your prompt with
 <code>
 real life photo
 </code>
 . If that's not enough, add
 <code>
 photo \(medium\)
 </code>
 and increase its strength until satisfied. You can usually go up to a crazy strength value like 5 or 6 without breaking the image.
 
 </li>
 <li>
 
 You can reduce the first number on the sigma curve to 0.95-0.99 to improve realism. This reduces saturation and adds a little noise, but makes the model less stable.
 
 </li>
 <li>
 
 You can remove NegPip fluff to improve anatomy (e.g. fingers) at the cost of some photographic texture.
 
 </li>
 <li>
 
 Newest workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/JjhHv266/image-photanima-turbo-v2-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Simple workflow with fewer custom nodes:
 <a href="https://i.ibb.co/p6t1YcL3/image-photanima-turbo-v2-0-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 Base/Non-Turbo:
 

<ul>
 <li>
 
 You can get a good image in 25 steps, but 40 is often better.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG between 3.5 to 4.
 
 </li>
 <li>
 
 Recommended fluff: "(photo \(medium\):1), real life, score_9, aesthetic"
 
 </li>
 <li>
 
 Recommended negative prompt: "toon \(style\), anime coloring, painting \(medium\), airbrushed, mutation, distortion, ai-assisted, glossy, shiny, shiny skin, worst quality, score_3, score_4"
 
 </li>
 <li>
 
 I have found it's helpful to decay conditioning strength from 2 to 1 over the first ~40% of steps. The stock workflow does this.
 
 </li>
 <li>
 
 Newest workflow optimized for realism:
 <a href="https://i.ibb.co/xtrs0F5Q/image-photanima-v2-1b.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>


<hr/>


<h2>
 
 🗺️ Roadmap
 
</h2>



 I'm pretty excited about the potential of Anima, but let's be clear: I'm not claiming that this checkpoint is a "ZIT killer." The correct model to compare this against is SDXL/IL - and I'm confident that Anima can dethrone it with enough community effort.


 Directions I'd like to explore next:

<ul>
 <li>
 
 (✅ Done in v2) There are a handful of Anima "detailer" LoRAs on Civitai. These are not intended for photography, but with enough block pruning, you never know. The right mix could go a long way.
 
 </li>
 <li>
 
 I suspect further increasing the dataset to ~3k images would help resolve remaining issues related to certain textures or model biases.
 
 </li>
 <li>
 
 (✅ Done in v2) I'm eagerly awaiting the release of Anima Turbo 1.0. The current Turbo solution is based on Preview3 and I think it's holding back this model's potential a little.
 
 </li>
 <li>
 
 I'm also looking forward to Anima support in OneTrainer. It will make trying experimental configs a lot less of a hassle compared to kohya-ss. For this v1 run, I stuck with safe values (prodigy, 1.0 LR, no fancy flags.)
 
 </li>
</ul>

v2.1 Turbo

( Not my Model )


 <a href="https://civitai.com/models/2645333/photanima?modelVersionId=3016852" rel="noopener noreferrer nofollow" target="_blank">
 https://civitai.com/models/2645333/photanima?modelVersionId=3016852
 </a>




 About this version


 Non-Turbo edition for merging or LoRA training. Check main model post for updated details on use.




 
 Photanima
 
 is an experimental finetune of
 
 Anima Base v1.0
 
 to see whether it is a viable architecture for photography. Spoiler alert: it totally is.


 Turbo LoRA baked in. If you're on a 30-series GPU, I recommend using this with the
 <a href="https://github.com/SparknightLLC/ComfyUI-INT8-Toolkit" rel="ugc" target="_blank">
 INT8 Toolkit
 </a>
 + INT8 Lazy Torch Compile node for wicked fast gen times. All demo images generated with that combo. These are raw outputs; no upscaling or post-processing.


 Most demo images contain workflows with custom sigma curve and ODE sampler. These both help significantly with realism. Standalone workflows provided further down this post.




 
 ❤️ If you enjoy Photanima, you can help offset the cost of training:
 


 <a href="https://ko-fi.com/liftweights" rel="ugc" target="_blank">
 Buy liftweights a Coffee
 </a>



<hr/>


<h2>
 
 🤓 Technical details
 
</h2>



 v2 is trained on ~2000 images for 45,000 steps. This is an expansion of my
 <a href="https://civitai.red/models/2068592/snakebite-2?modelVersionId=2426502" rel="ugc" target="_blank">
 Snakebite 2.3
 </a>
 dataset with around 700 new images and captions reworked for Anima. Training took approximately 48 hours on a Geforce 3090.




 
 Pros:
 

<ul>
 <li>
 
 Extremely fast.
 
 </li>
 <li>
 
 Extremely good prompt adherence.
 
 </li>
 <li>
 
 Anatomy is pretty stable. If it screws something up, changing your steps by +1/-1 usually fixes it.
 
 </li>
 <li>
 
 Supports up to nearly 2MP with little-to-no distortions.
 
 </li>
</ul>



 At first, I noticed that Photanima's style was inconsistent - it had a tendency to regress toward a cartoony/CGI look as my prompts became more complex. I was able to mostly overcome this by splitting Photanima into constituent content, style-early, and style-late blocks, then boosted the style blocks well past a strength of 1.


 "Style-late" maps to blocks 7, 8, and 9 - these do alter composition to a degree, so we can't boost them as hard as "style-early."


 Images are pretty consistent now, but there are some notable drawbacks.




 
 Cons in v2:
 

<ul>
 <li>
 
 It loses a little knowledge of certain artistic terms like
 <code>
 silhouette
 </code>
 .
 
 </li>
 <li>
 
 Microdetail quality is somewhere between SDXL and ZIT. Honestly, it's really good for a 2B model. Two-step upscaling with Anima doesn't help much, but I'm sure the results would be amazing if you sent a Photanima image to a different model for refinement. Or if that's too much work: just add a little film grain. It does wonders and requires no extra VRAM.
 
 </li>
 <li>
 
 Text capabilities are not as good as those of base Anima. Anything beyond 3 or 4 words is likely going to require numerous re-rolls. This is at least partly due to the Turbo LoRA.
 
 </li>
 <li>
 
 Excessive fluff tags like
 <code>
 masterpiece, absurdres, hyperreal
 </code>
 tend to fry the image. The model is photographic and highly aesthetic by default, so there's no need to drive it harder in that direction.
 
 </li>
</ul>




<hr/>


<h2>
 
 🛠️ Recommended Settings (for latest versions)
 
</h2>



 
 Turbo:
 

<ul>
 <li>
 
 6-8 steps. Images often look best at 6, but anatomy is more stable at 8-10, especially with complex prompts.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG exactly 1.
 
 </li>
 <li>
 
 Preferred resolution: 1040x1520 or 832x1216.
 
 </li>
 <li>
 
 For maximum realism, begin your prompt with
 <code>
 real life photo
 </code>
 . If that's not enough, add
 <code>
 photo \(medium\)
 </code>
 and increase its strength until satisfied. You can usually go up to a crazy strength value like 5 or 6 without breaking the image.
 
 </li>
 <li>
 
 You can reduce the first number on the sigma curve to 0.95-0.99 to improve realism. This reduces saturation and adds a little noise, but makes the model less stable.
 
 </li>
 <li>
 
 You can remove NegPip fluff to improve anatomy (e.g. fingers) at the cost of some photographic texture.
 
 </li>
 <li>
 
 Newest workflow optimized for realism (recommended):
 <a href="https://i.ibb.co/JjhHv266/image-photanima-turbo-v2-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
 <li>
 
 Simple workflow with fewer custom nodes:
 <a href="https://i.ibb.co/p6t1YcL3/image-photanima-turbo-v2-0-1.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>



 
 Base/Non-Turbo:
 

<ul>
 <li>
 
 You can get a good image in 25 steps, but 40 is often better.
 
 </li>
 <li>
 
 er_sde sampler on "ODE" mode.
 
 </li>
 <li>
 
 Custom sigma curve or simple scheduler: "1, 0.94, 0.9, 0.825, 0.6, 0.5, 0.3, 0.29, 0.2, 0.0"
 
 </li>
 <li>
 
 CFG between 3.5 to 4.
 
 </li>
 <li>
 
 Recommended fluff: "(photo \(medium\):1), real life, score_9, aesthetic"
 
 </li>
 <li>
 
 Recommended negative prompt: "toon \(style\), anime coloring, painting \(medium\), airbrushed, mutation, distortion, ai-assisted, glossy, shiny, shiny skin, worst quality, score_3, score_4"
 
 </li>
 <li>
 
 I have found it's helpful to decay conditioning strength from 2 to 1 over the first ~40% of steps. The stock workflow does this.
 
 </li>
 <li>
 
 Newest workflow optimized for realism:
 <a href="https://i.ibb.co/xtrs0F5Q/image-photanima-v2-1b.png" rel="ugc" target="_blank">
 Download
 </a>
 
 </li>
</ul>


<hr/>


<h2>
 
 🗺️ Roadmap
 
</h2>



 I'm pretty excited about the potential of Anima, but let's be clear: I'm not claiming that this checkpoint is a "ZIT killer." The correct model to compare this against is SDXL/IL - and I'm confident that Anima can dethrone it with enough community effort.


 Directions I'd like to explore next:

<ul>
 <li>
 
 (✅ Done in v2) There are a handful of Anima "detailer" LoRAs on Civitai. These are not intended for photography, but with enough block pruning, you never know. The right mix could go a long way.
 
 </li>
 <li>
 
 I suspect further increasing the dataset to ~3k images would help resolve remaining issues related to certain textures or model biases.
 
 </li>
 <li>
 
 (✅ Done in v2) I'm eagerly awaiting the release of Anima Turbo 1.0. The current Turbo solution is based on Preview3 and I think it's holding back this model's potential a little.
 
 </li>
 <li>
 
 I'm also looking forward to Anima support in OneTrainer. It will make trying experimental configs a lot less of a hassle compared to kohya-ss. For this v1 run, I stuck with safe values (prodigy, 1.0 LR, no fancy flags.)
 
 </li>
</ul>