Instructions to use darask0/anima-distill-loras with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use darask0/anima-distill-loras with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("circlestone-labs/Anima", dtype=torch.bfloat16, device_map="cuda") pipe.load_lora_weights("darask0/anima-distill-loras") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
Upload pcm/README.md with huggingface_hub
Browse files- pcm/README.md +50 -16
pcm/README.md
CHANGED
|
@@ -13,20 +13,30 @@ library_name: diffusers
|
|
| 13 |
|
| 14 |
# Anima v1.0 — PCM 4-step Distillation LoRA
|
| 15 |
|
| 16 |
-
|
|
|
|
| 17 |
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
## Files
|
| 21 |
|
| 22 |
| File | Format | Use |
|
| 23 |
|---|---|---|
|
| 24 |
-
| `pcm_final_peft.safetensors` | PEFT (diffusers) |
|
| 25 |
-
| `pcm_final_comfy.safetensors` | ComfyUI LoRA |
|
| 26 |
|
| 27 |
両者は同一の重み、フォーマット変換のみの違い。
|
| 28 |
|
| 29 |
-
## Usage (ComfyUI)
|
| 30 |
|
| 31 |
1. `pcm_final_comfy.safetensors` を `ComfyUI/models/loras/` に配置
|
| 32 |
2. Anima v1.0 base workflow に `LoraLoaderModelOnly` を挿入、`strength_model: 1.0`
|
|
@@ -35,16 +45,14 @@ Phased Consistency Model (PCM) 蒸留 LoRA for [Anima v1.0 base](https://hugging
|
|
| 35 |
- **cfg: 1.0**
|
| 36 |
- sampler / scheduler: 下記サンプル比較を参照
|
| 37 |
|
| 38 |
-
## Recommended Sampler
|
| 39 |
-
|
| 40 |
-
訓練後 step 3500-4000 で複数 sampler を比較済み:
|
| 41 |
|
| 42 |
| Sampler | Scheduler | 特徴 |
|
| 43 |
|---|---|---|
|
| 44 |
| `er_sde` | `simple` | やや softer、構図上半身寄り、生成 8.8s |
|
| 45 |
| `res_multistep` | `beta` | anime style 強め、全身描画、生成 11.2s |
|
| 46 |
|
| 47 |
-
両者とも 4-step CFG=1.0 で
|
| 48 |
|
| 49 |
### サンプル比較
|
| 50 |
|
|
@@ -74,9 +82,10 @@ Phased Consistency Model (PCM) 蒸留 LoRA for [Anima v1.0 base](https://hugging
|
|
| 74 |
| Loss | Pseudo-Huber |
|
| 75 |
| Optimizer | AdamW (lr 5e-6, wd 0.01) |
|
| 76 |
| Grad clip | 1.0 |
|
|
|
|
| 77 |
| Hardware | NVIDIA B200 (1 GPU) |
|
| 78 |
| Training time | ~3.4 hours |
|
| 79 |
-
| Cost | ~$
|
| 80 |
|
| 81 |
### Dataset
|
| 82 |
|
|
@@ -85,24 +94,49 @@ Anima v1.0 self-distillation cache (5000 caption × teacher x0 + crossattn embed
|
|
| 85 |
- Caption pool: 10 prompts × 500 seeds, anime characters + scene
|
| 86 |
- Precompute cost: ~$7-8 (B200, ~86 min)
|
| 87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
## Limitations
|
| 89 |
|
| 90 |
-
-
|
| 91 |
-
-
|
| 92 |
-
-
|
| 93 |
-
-
|
| 94 |
|
| 95 |
## License
|
| 96 |
|
| 97 |
Apache-2.0
|
| 98 |
|
| 99 |
-
Base model (Anima v1.0) のライセンスも併せて確認のこと:
|
|
|
|
| 100 |
|
| 101 |
## Citation
|
| 102 |
|
| 103 |
PCM 元論文:
|
| 104 |
|
| 105 |
-
```
|
| 106 |
@article{wang2024phased,
|
| 107 |
title={Phased Consistency Model},
|
| 108 |
author={Wang, Fu-Yun and Huang, Zhaoyang and Bei, Bin and Shi, Xiaoyu and Liu, Xinyu and Tian, Yang and Yang, Yang and Li, Hongsheng},
|
|
|
|
| 13 |
|
| 14 |
# Anima v1.0 — PCM 4-step Distillation LoRA
|
| 15 |
|
| 16 |
+
⚠️ **非推奨 / Not Recommended**: 本 LoRA は生成品質が低く、実用には Civitai Anima Turbo を推奨。
|
| 17 |
+
研究目的・失敗事例の参照用として公開。
|
| 18 |
|
| 19 |
+
Phased Consistency Model (PCM) 蒸留 LoRA for [Anima v1.0 base](https://huggingface.co/circlestone-labs/Anima)。
|
| 20 |
+
4-step / CFG=1.0 で生成可能だが、**視覚品質が公式 [Civitai Anima Turbo](https://civitai.com/models/2560840)
|
| 21 |
+
や同 collection の DMDX に劣る**。
|
| 22 |
+
|
| 23 |
+
## なぜ非推奨か
|
| 24 |
+
|
| 25 |
+
- 5000 step 訓練を **数値的に完走** (loss 健全、divergence なし) したが、生成画像の品質が低い
|
| 26 |
+
- 実用は **公式 Civitai Anima Turbo を直接使用** が確実
|
| 27 |
+
|
| 28 |
+
詳細な失敗分析は [docs/pcm.md](https://huggingface.co/darask0/rapid-anima/blob/main/docs/pcm.md) 参照。
|
| 29 |
|
| 30 |
## Files
|
| 31 |
|
| 32 |
| File | Format | Use |
|
| 33 |
|---|---|---|
|
| 34 |
+
| `pcm_final_peft.safetensors` | PEFT (diffusers) | 研究用 |
|
| 35 |
+
| `pcm_final_comfy.safetensors` | ComfyUI LoRA | 研究用 |
|
| 36 |
|
| 37 |
両者は同一の重み、フォーマット変換のみの違い。
|
| 38 |
|
| 39 |
+
## Usage (ComfyUI、研究用)
|
| 40 |
|
| 41 |
1. `pcm_final_comfy.safetensors` を `ComfyUI/models/loras/` に配置
|
| 42 |
2. Anima v1.0 base workflow に `LoraLoaderModelOnly` を挿入、`strength_model: 1.0`
|
|
|
|
| 45 |
- **cfg: 1.0**
|
| 46 |
- sampler / scheduler: 下記サンプル比較を参照
|
| 47 |
|
| 48 |
+
## Recommended Sampler (どちらも品質不足だが)
|
|
|
|
|
|
|
| 49 |
|
| 50 |
| Sampler | Scheduler | 特徴 |
|
| 51 |
|---|---|---|
|
| 52 |
| `er_sde` | `simple` | やや softer、構図上半身寄り、生成 8.8s |
|
| 53 |
| `res_multistep` | `beta` | anime style 強め、全身描画、生成 11.2s |
|
| 54 |
|
| 55 |
+
両者とも 4-step CFG=1.0 で動作するが、品質的に Civitai Turbo / DMDX より劣る。
|
| 56 |
|
| 57 |
### サンプル比較
|
| 58 |
|
|
|
|
| 82 |
| Loss | Pseudo-Huber |
|
| 83 |
| Optimizer | AdamW (lr 5e-6, wd 0.01) |
|
| 84 |
| Grad clip | 1.0 |
|
| 85 |
+
| Warm-start | **無し** (cold-start) |
|
| 86 |
| Hardware | NVIDIA B200 (1 GPU) |
|
| 87 |
| Training time | ~3.4 hours |
|
| 88 |
+
| Cost | ~$22 (Modal) |
|
| 89 |
|
| 90 |
### Dataset
|
| 91 |
|
|
|
|
| 94 |
- Caption pool: 10 prompts × 500 seeds, anime characters + scene
|
| 95 |
- Precompute cost: ~$7-8 (B200, ~86 min)
|
| 96 |
|
| 97 |
+
## 失敗分析 (Lessons Learned)
|
| 98 |
+
|
| 99 |
+
### 数値上は成功した
|
| 100 |
+
- loss 0.001-0.007 で安定収束
|
| 101 |
+
- mean collapse / NaN / OOM 全部なし
|
| 102 |
+
- training pipeline は完璧に動作
|
| 103 |
+
|
| 104 |
+
### しかし生成品質は低い
|
| 105 |
+
推測される原因:
|
| 106 |
+
|
| 107 |
+
1. **cold-start strategy が Anima 系では機能しない可能性**
|
| 108 |
+
- 公式 Turbo は warm-start が事実上必須
|
| 109 |
+
- 5000 sample dataset は cold-start には不十分かも
|
| 110 |
+
2. **phase consistency loss だけでは style anchor が弱い**
|
| 111 |
+
- DMD2 系の adversarial signal や LADD の Smooth-L1 recon があれば改善可能性
|
| 112 |
+
3. **CFG-augmentation embed が不完全**
|
| 113 |
+
- w∈[4.0, 5.0] embed が teacher CFG=4.5 の効果を完全再現していない
|
| 114 |
+
|
| 115 |
+
### 改善方向 (将来検討)
|
| 116 |
+
- warm-start (Civitai Turbo) を導入して semi-fine-tune に
|
| 117 |
+
- データセット規模拡大 (5000 → 20000+)
|
| 118 |
+
- adversarial loss (LADD style) との組合せ
|
| 119 |
+
- 他手法 (DMD2、Hyper-SD) に切り替え
|
| 120 |
+
|
| 121 |
## Limitations
|
| 122 |
|
| 123 |
+
- ❌ **生成品質が公式 Turbo / DMDX より劣る** — 主要な制約
|
| 124 |
+
- Base model 固定 (Anima v1.0 以外では style drift)
|
| 125 |
+
- 4-step ハード制約 (2 step / 8 step では同等品質保証なし)
|
| 126 |
+
- CFG=1.0 推奨 (CFG>1.0 は過剰)
|
| 127 |
|
| 128 |
## License
|
| 129 |
|
| 130 |
Apache-2.0
|
| 131 |
|
| 132 |
+
Base model (Anima v1.0) のライセンスも併せて確認のこと:
|
| 133 |
+
[circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima)
|
| 134 |
|
| 135 |
## Citation
|
| 136 |
|
| 137 |
PCM 元論文:
|
| 138 |
|
| 139 |
+
```bibtex
|
| 140 |
@article{wang2024phased,
|
| 141 |
title={Phased Consistency Model},
|
| 142 |
author={Wang, Fu-Yun and Huang, Zhaoyang and Bei, Bin and Shi, Xiaoyu and Liu, Xinyu and Tian, Yang and Yang, Yang and Li, Hongsheng},
|