Replace the head with realr_1119_2200 (original-image target + LPIPS at the base points, turbo point on v4_longer); previous head stays in old/
Browse files- README.md +8 -2
- anima_pixel_head.safetensors +2 -2
README.md
CHANGED
|
@@ -23,7 +23,7 @@ writes RGB pixels directly, in place of the VAE decode. The head and its trainin
|
|
| 23 |
| Output | a 16×16 RGB patch per token, in [-1, 1] |
|
| 24 |
| Architecture | linear projection of the feature + linear embedding of the x_t patch, 2 ConvNeXt, one 2× pixel-shuffle upsample, 1 ConvNeXt, linear to 8×8×3 |
|
| 25 |
| DiT | `anima-base-v1.0` (28 blocks) |
|
| 26 |
-
| Target | VAE(x̂0) of the same forward, x̂0 = x_t − σ·v |
|
| 27 |
|
| 28 |
|
| 29 |
## Compatibility
|
|
@@ -34,10 +34,16 @@ Read the head only where it was trained:
|
|
| 34 |
Spectrum included (its last step is an actual forward).
|
| 35 |
- **Turbo**: the 4-step turbo LoRA of
|
| 36 |
[anima-turbo-4step](https://huggingface.co/sorryhyun/anima-turbo-4step), Euler 4 steps,
|
| 37 |
-
shift 3, CFG 1 (last-step σ 0.5).
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
## Benchmark
|
| 40 |
|
|
|
|
|
|
|
|
|
|
| 41 |
### In ComfyUI
|
| 42 |
|
| 43 |
Turbo 4 steps, CFG 1, RTX 5070 Ti, 16 prompts at ~1 MP, one seed
|
|
|
|
| 23 |
| Output | a 16×16 RGB patch per token, in [-1, 1] |
|
| 24 |
| Architecture | linear projection of the feature + linear embedding of the x_t patch, 2 ConvNeXt, one 2× pixel-shuffle upsample, 1 ConvNeXt, linear to 8×8×3 |
|
| 25 |
| DiT | `anima-base-v1.0` (28 blocks) |
|
| 26 |
+
| Target | base points: the original image, with LPIPS (weight 0.05); turbo point: VAE(x̂0) of the same forward, x̂0 = x_t − σ·v |
|
| 27 |
|
| 28 |
|
| 29 |
## Compatibility
|
|
|
|
| 34 |
Spectrum included (its last step is an actual forward).
|
| 35 |
- **Turbo**: the 4-step turbo LoRA of
|
| 36 |
[anima-turbo-4step](https://huggingface.co/sorryhyun/anima-turbo-4step), Euler 4 steps,
|
| 37 |
+
shift 3, CFG 1 (last-step σ 0.5). Trained on `anima_turbo_4step_v4_longer.safetensors`.
|
| 38 |
+
|
| 39 |
+
The previous head is `old/anima_pixel_head.safetensors` (VAE(x̂0) target at every point, turbo
|
| 40 |
+
point on that repo's `old/anima_turbo_4step.safetensors`).
|
| 41 |
|
| 42 |
## Benchmark
|
| 43 |
|
| 44 |
+
The benchmarks, samples and held-out reads below were measured on the previous head
|
| 45 |
+
(`old/anima_pixel_head.safetensors`).
|
| 46 |
+
|
| 47 |
### In ComfyUI
|
| 48 |
|
| 49 |
Turbo 4 steps, CFG 1, RTX 5070 Ti, 16 prompts at ~1 MP, one seed
|
anima_pixel_head.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2b9d24cfb84a38aa9c764f21ed47496f6dd2ccb2ceff063f467ed3d1a7f4a5a2
|
| 3 |
+
size 99442192
|