この派生モデルは元々のものを保持しようとしていますが、「realistic」および「photorealistic」のタグを少しうまく機能させるようにしています。これらが省略されると、デフォルトでアニメ風になります。完璧なフォトリアリズムではありませんが、2.5Dからセミリアルな感じで、時折非常にフォトリアルな生成が見られることもあります。

The amazing Animagine XL 4.0 knows many characters, poses, camera angles, expressions and emotions. It is quite "magical" compared to other models.

This derivative model tries to preserve the original, but makes the "photorealistic" tag work a little better. If realistic and photorealistic are omitted, it defaults to anime. It is not perfect photorealism, but, it is 2.5D to semi-real, with occasional generations looking very photoreal indeed.

If you like my model, please give it a 👍 and write few words of review, it really motivates me!

v4: steps: 25 cfg: 4.0 sampler: euler_a scheduler: auto/normal.
v3: steps: 20 cfg: 2.0 sampler: dpmpp_2m scheduler: beta (reduces noise!)

size: 1024 x 1408 (1536 often creates long body or long limbs)

negative: lowres, error, cropped, worst quality, low quality, low score, bad score, average score, (blurry, nsfw, explicit, nude)

realism trigger words: sunlight, shadow, photorealistic. Do not increase strength in ComfyUI with a short prompt, (photorealistic:1.5) will destroy the quality.

For dark images, lower cfg and use the "dramatic lighting" lora. These keywords help a little: (chiaroscuro, darkness, dark room, shadows, dusk, night:1.2)

Workflow: I have uploaded a zip of the showcase images in PNG format for you to drag in ComfyUI workflows. It's called "Training Data" as that was the only option.

Animagine recommends not to start with "masterpiece, high score, great score, absurdres realistic, photorealistic" but I find it works fine "(photorealistic), masterpiece, high score, great score, absurdres"

original prompting guide: https://cagliostrolab.net/posts/optimizing-animagine-xl-40-in-depth-guideline-and-update

Version History

v1 - quite authentic, quite realistic, but quite noisy except with specific samplers. Not recommended.

v2 - improved photorealism but further from the Anim4gine base model. Not really recommended.

v3 - I focused on improving the prompt following from the base model, face expressions and good photo-reality. Works with wider range of samplers, steps and cfg. Recommended.

v4 - An improvement in anatomy, realism, highlights and shadows. It also follows closely to Animagine XL 4.0. Use "safe" in the positive and (nsfw, explicit) in the negative.

Please share any questions, comments or better configurations, merges or loras you have found in the discussion below!

full size image link

License

This model adopts the original CreativeML Open RAIL++-M License from Stability AI without any modifications or additional restrictions. The license terms remain exactly as specified in the original SDXL license, which includes:

✅ Permitted: Commercial use, modifications, distributions, private use
❌ Prohibited: Illegal activities, harmful content generation, discrimination, exploitation
⚠️ Requirements: Include license copy, state changes, preserve notices
📝 Warranty: Provided "AS IS" without warranties

Please refer to the original SDXL license for the complete and authoritative terms and conditions.