Moss anime UWU alt turbo

STEPS 8-12

sampler

Experiment with it; the model turned out to be quite flexible.

scheduler

bong_tangent\SGM uniform\simple\beta 57

CFG 1-1.1

I think the tuning tool turned out well.

MOSS Block Strength Editor (tune checkpoint) - v1.0 | MOSS Image Other | Civitai

Moss Yuni anime UWU base

This is a completely different model with reconfigured weight blocks. It handles details better and produces superior colors compared to the base 'moss' and 'SFT' models. However, it doesn't work very well with Turbo LoRAs. I have no desire to reconfigure it for Turbo, as I already have a Turbo version that works just fine.

but it also contains a dataset that was used to create the turbo version

INT8 will be with the assembled qkv (block atention) BF16 q+k+v (block atention)

STEPS

25-50 (optimal 25-30)

sampler

dpmpp 2m_SDE\ dpmpp SDE

scheduler

SGM uniform\simple

CFG 3-5

For those interested, here is the structure of MOSS SFT

MOSS checkpoint structure report

===========================================================

Input file: moss_image_sft_bf16.safetensors

Total keys: 379

Top key prefixes (first two components):

28 text_embedder.blocks.

8 layers.0.

8 layers.1.

8 layers.10.

8 layers.11.

8 layers.12.

8 layers.13.

8 layers.14.

8 layers.15.

8 layers.16.

8 layers.17.

8 layers.18.

8 layers.19.

8 layers.2.

8 layers.20.

8 layers.21.

8 layers.22.

8 layers.23.

8 layers.24.

8 layers.25.

8 layers.26.

8 layers.27.

8 layers.28.

8 layers.29.

8 layers.3.

8 layers.30.

8 layers.31.

8 layers.32.

8 layers.33.

8 layers.34.

8 layers.35.

8 layers.36.

8 layers.37.

8 layers.38.

8 layers.39.

8 layers.4.

8 layers.40.

8 layers.41.

8 layers.5.

8 layers.6.

MOSS linear prefix counts:

168 ''

Raw matched MOSS-like linear tensors:

total: 168

Raw part counts:

qkv : 42

q : 0

k : 0

v : 0

o : 42

ffn_up_gate : 42

ffn_down : 42

Detected blocks:

total blocks: 42

block numbers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41]

Fusion estimate:

blocks with fused qkv key: 42

blocks with complete separate q/k/v: 0

blocks that can become qkv: 42

blocks with o: 42

blocks with ffn_up_gate: 42

blocks with ffn_down: 42

estimated linears after q/k/v fusion: 168

Missing expected parts by block:

missing qkv or complete q/k/v: []

missing o: []

missing ffn_up_gate: []

missing ffn_down: []

Blocks with incomplete separate q/k/v:

[]

Other / non-linear keys:

Possible adapter-like keys:

none detected by simple markers

First 150 other keys:

final_linear_video.weight

final_norm_video.weight

layers.0.k_norm.weight

layers.0.norm1.weight

layers.0.norm2.weight

layers.0.q_norm.weight

layers.1.k_norm.weight

layers.1.norm1.weight

layers.1.norm2.weight

layers.1.q_norm.weight

layers.10.k_norm.weight

layers.10.norm1.weight

layers.10.norm2.weight

layers.10.q_norm.weight

layers.11.k_norm.weight

layers.11.norm1.weight

layers.11.norm2.weight

layers.11.q_norm.weight

layers.12.k_norm.weight

layers.12.norm1.weight

layers.12.norm2.weight

layers.12.q_norm.weight

layers.13.k_norm.weight

layers.13.norm1.weight

layers.13.norm2.weight

layers.13.q_norm.weight

layers.14.k_norm.weight

layers.14.norm1.weight

layers.14.norm2.weight

layers.14.q_norm.weight

layers.15.k_norm.weight

layers.15.norm1.weight

layers.15.norm2.weight

layers.15.q_norm.weight

layers.16.k_norm.weight

layers.16.norm1.weight

layers.16.norm2.weight

layers.16.q_norm.weight

layers.17.k_norm.weight

layers.17.norm1.weight

layers.17.norm2.weight

layers.17.q_norm.weight

layers.18.k_norm.weight

layers.18.norm1.weight

layers.18.norm2.weight

layers.18.q_norm.weight

layers.19.k_norm.weight

layers.19.norm1.weight

layers.19.norm2.weight

layers.19.q_norm.weight

layers.2.k_norm.weight

layers.2.norm1.weight

layers.2.norm2.weight

layers.2.q_norm.weight

layers.20.k_norm.weight

layers.20.norm1.weight

layers.20.norm2.weight

layers.20.q_norm.weight

layers.21.k_norm.weight

layers.21.norm1.weight

layers.21.norm2.weight

layers.21.q_norm.weight

layers.22.k_norm.weight

layers.22.norm1.weight

layers.22.norm2.weight

layers.22.q_norm.weight

layers.23.k_norm.weight

layers.23.norm1.weight

layers.23.norm2.weight

layers.23.q_norm.weight

layers.24.k_norm.weight

layers.24.norm1.weight

layers.24.norm2.weight

layers.24.q_norm.weight

layers.25.k_norm.weight

layers.25.norm1.weight

layers.25.norm2.weight

layers.25.q_norm.weight

layers.26.k_norm.weight

layers.26.norm1.weight

layers.26.norm2.weight

layers.26.q_norm.weight

layers.27.k_norm.weight

layers.27.norm1.weight

layers.27.norm2.weight

layers.27.q_norm.weight

layers.28.k_norm.weight

layers.28.norm1.weight

layers.28.norm2.weight

layers.28.q_norm.weight

layers.29.k_norm.weight

layers.29.norm1.weight

layers.29.norm2.weight

layers.29.q_norm.weight

layers.3.k_norm.weight

layers.3.norm1.weight

layers.3.norm2.weight

layers.3.q_norm.weight

layers.30.k_norm.weight

layers.30.norm1.weight

layers.30.norm2.weight

layers.30.q_norm.weight

layers.31.k_norm.weight

layers.31.norm1.weight

layers.31.norm2.weight

layers.31.q_norm.weight

layers.32.k_norm.weight

layers.32.norm1.weight

layers.32.norm2.weight

layers.32.q_norm.weight

layers.33.k_norm.weight

layers.33.norm1.weight

layers.33.norm2.weight

layers.33.q_norm.weight

layers.34.k_norm.weight

layers.34.norm1.weight

layers.34.norm2.weight

layers.34.q_norm.weight

layers.35.k_norm.weight

layers.35.norm1.weight

layers.35.norm2.weight

layers.35.q_norm.weight

layers.36.k_norm.weight

layers.36.norm1.weight

layers.36.norm2.weight

layers.36.q_norm.weight

layers.37.k_norm.weight

layers.37.norm1.weight

layers.37.norm2.weight

layers.37.q_norm.weight

layers.38.k_norm.weight

layers.38.norm1.weight

layers.38.norm2.weight

layers.38.q_norm.weight

layers.39.k_norm.weight

layers.39.norm1.weight

layers.39.norm2.weight

layers.39.q_norm.weight

layers.4.k_norm.weight

layers.4.norm1.weight

layers.4.norm2.weight

layers.4.q_norm.weight

layers.40.k_norm.weight

layers.40.norm1.weight

layers.40.norm2.weight

layers.40.q_norm.weight

layers.41.k_norm.weight

layers.41.norm1.weight

layers.41.norm2.weight

layers.41.q_norm.weight

... 61 more

structure of MOSS

Input file: moss_image_base_bf16.safetensors

Total keys: 379

Top key prefixes (first two components):

28 text_embedder.blocks.

8 layers.0.

8 layers.1.

8 layers.10.

8 layers.11.

8 layers.12.

8 layers.13.

8 layers.14.

8 layers.15.

8 layers.16.

8 layers.17.

8 layers.18.

8 layers.19.

8 layers.2.

8 layers.20.

8 layers.21.

8 layers.22.

8 layers.23.

8 layers.24.

8 layers.25.

8 layers.26.

8 layers.27.

8 layers.28.

8 layers.29.

8 layers.3.

8 layers.30.

8 layers.31.

8 layers.32.

8 layers.33.

8 layers.34.

8 layers.35.

8 layers.36.

8 layers.37.

8 layers.38.

8 layers.39.

8 layers.4.

8 layers.40.

8 layers.41.

8 layers.5.

8 layers.6.

MOSS linear prefix counts:

168 ''

Raw matched MOSS-like linear tensors:

total: 168

Raw part counts:

qkv : 42

q : 0

k : 0

v : 0

o : 42

ffn_up_gate : 42

ffn_down : 42

Detected blocks:

total blocks: 42

block numbers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41]

Fusion estimate:

blocks with fused qkv key: 42

blocks with complete separate q/k/v: 0

blocks that can become qkv: 42

blocks with o: 42

blocks with ffn_up_gate: 42

blocks with ffn_down: 42

estimated linears after q/k/v fusion: 168

Missing expected parts by block:

missing qkv or complete q/k/v: []

missing o: []

missing ffn_up_gate: []

missing ffn_down: []

Blocks with incomplete separate q/k/v:

[]

Other / non-linear keys:

Possible adapter-like keys:

none detected by simple markers

First 150 other keys:

final_linear_video.weight

final_norm_video.weight

layers.0.k_norm.weight

layers.0.norm1.weight

layers.0.norm2.weight

layers.0.q_norm.weight

layers.1.k_norm.weight

layers.1.norm1.weight

layers.1.norm2.weight

layers.1.q_norm.weight

layers.10.k_norm.weight

layers.10.norm1.weight

layers.10.norm2.weight

layers.10.q_norm.weight

layers.11.k_norm.weight

layers.11.norm1.weight

layers.11.norm2.weight

layers.11.q_norm.weight

layers.12.k_norm.weight

layers.12.norm1.weight

layers.12.norm2.weight

layers.12.q_norm.weight

layers.13.k_norm.weight

layers.13.norm1.weight

layers.13.norm2.weight

layers.13.q_norm.weight

layers.14.k_norm.weight

layers.14.norm1.weight

layers.14.norm2.weight

layers.14.q_norm.weight

layers.15.k_norm.weight

layers.15.norm1.weight

layers.15.norm2.weight

layers.15.q_norm.weight

layers.16.k_norm.weight

layers.16.norm1.weight

layers.16.norm2.weight

layers.16.q_norm.weight

layers.17.k_norm.weight

layers.17.norm1.weight

layers.17.norm2.weight

layers.17.q_norm.weight

layers.18.k_norm.weight

layers.18.norm1.weight

layers.18.norm2.weight

layers.18.q_norm.weight

layers.19.k_norm.weight

layers.19.norm1.weight

layers.19.norm2.weight

layers.19.q_norm.weight

layers.2.k_norm.weight

layers.2.norm1.weight

layers.2.norm2.weight

layers.2.q_norm.weight

layers.20.k_norm.weight

layers.20.norm1.weight

layers.20.norm2.weight

layers.20.q_norm.weight

layers.21.k_norm.weight

layers.21.norm1.weight

layers.21.norm2.weight

layers.21.q_norm.weight

layers.22.k_norm.weight

layers.22.norm1.weight

layers.22.norm2.weight

layers.22.q_norm.weight

layers.23.k_norm.weight

layers.23.norm1.weight

layers.23.norm2.weight

layers.23.q_norm.weight

layers.24.k_norm.weight

layers.24.norm1.weight

layers.24.norm2.weight

layers.24.q_norm.weight

layers.25.k_norm.weight

layers.25.norm1.weight

layers.25.norm2.weight

layers.25.q_norm.weight

layers.26.k_norm.weight

layers.26.norm1.weight

layers.26.norm2.weight

layers.26.q_norm.weight

layers.27.k_norm.weight

layers.27.norm1.weight

layers.27.norm2.weight

layers.27.q_norm.weight

layers.28.k_norm.weight

layers.28.norm1.weight

layers.28.norm2.weight

layers.28.q_norm.weight

layers.29.k_norm.weight

layers.29.norm1.weight

layers.29.norm2.weight

layers.29.q_norm.weight

layers.3.k_norm.weight

layers.3.norm1.weight

layers.3.norm2.weight

layers.3.q_norm.weight

layers.30.k_norm.weight

layers.30.norm1.weight

layers.30.norm2.weight

layers.30.q_norm.weight

layers.31.k_norm.weight

layers.31.norm1.weight

layers.31.norm2.weight

layers.31.q_norm.weight

layers.32.k_norm.weight

layers.32.norm1.weight

layers.32.norm2.weight

layers.32.q_norm.weight

layers.33.k_norm.weight

layers.33.norm1.weight

layers.33.norm2.weight

layers.33.q_norm.weight

layers.34.k_norm.weight

layers.34.norm1.weight

layers.34.norm2.weight

layers.34.q_norm.weight

layers.35.k_norm.weight

layers.35.norm1.weight

layers.35.norm2.weight

layers.35.q_norm.weight

layers.36.k_norm.weight

layers.36.norm1.weight

layers.36.norm2.weight

layers.36.q_norm.weight

layers.37.k_norm.weight

layers.37.norm1.weight

layers.37.norm2.weight

layers.37.q_norm.weight

layers.38.k_norm.weight

layers.38.norm1.weight

layers.38.norm2.weight

layers.38.q_norm.weight

layers.39.k_norm.weight

layers.39.norm1.weight

layers.39.norm2.weight

layers.39.q_norm.weight

layers.4.k_norm.weight

layers.4.norm1.weight

layers.4.norm2.weight

layers.4.q_norm.weight

layers.40.k_norm.weight

layers.40.norm1.weight

layers.40.norm2.weight

layers.40.q_norm.weight

layers.41.k_norm.weight

layers.41.norm1.weight

layers.41.norm2.weight

layers.41.q_norm.weight

... 61 more

As for me, it’s the following:

MOSS checkpoint structure report

===========================================================

Input file: moss anime UwU base .safetensors

Total keys: 462

Top key prefixes (first two components):

462 model.diffusion_model.

MOSS linear prefix counts:

252 'model.diffusion_model.'

Raw matched MOSS-like linear tensors:

total: 252

Raw part counts:

qkv : 0

q : 42

k : 42

v : 42

o : 42

ffn_up_gate : 42

ffn_down : 42

Detected blocks:

total blocks: 42

block numbers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41]

Fusion estimate:

blocks with fused qkv key: 0

blocks with complete separate q/k/v: 42

blocks that can become qkv: 42

blocks with o: 42

blocks with ffn_up_gate: 42

blocks with ffn_down: 42

estimated linears after q/k/v fusion: 168

Missing expected parts by block:

missing qkv or complete q/k/v: []

missing o: []

missing ffn_up_gate: []

missing ffn_down: []

Blocks with incomplete separate q/k/v:

[]

Other / non-linear keys:

Possible adapter-like keys:

none detected by simple markers

First 150 other keys:

model.diffusion_model.final_linear_video.weight

model.diffusion_model.final_norm_video.weight

model.diffusion_model.layers.0.k_norm.weight

model.diffusion_model.layers.0.norm1.weight

model.diffusion_model.layers.0.norm2.weight

model.diffusion_model.layers.0.q_norm.weight

model.diffusion_model.layers.1.k_norm.weight

model.diffusion_model.layers.1.norm1.weight

model.diffusion_model.layers.1.norm2.weight

model.diffusion_model.layers.1.q_norm.weight

model.diffusion_model.layers.10.k_norm.weight

model.diffusion_model.layers.10.norm1.weight

model.diffusion_model.layers.10.norm2.weight

model.diffusion_model.layers.10.q_norm.weight

model.diffusion_model.layers.11.k_norm.weight

model.diffusion_model.layers.11.norm1.weight

model.diffusion_model.layers.11.norm2.weight

model.diffusion_model.layers.11.q_norm.weight

model.diffusion_model.layers.12.k_norm.weight

model.diffusion_model.layers.12.norm1.weight

model.diffusion_model.layers.12.norm2.weight

model.diffusion_model.layers.12.q_norm.weight

model.diffusion_model.layers.13.k_norm.weight

model.diffusion_model.layers.13.norm1.weight

model.diffusion_model.layers.13.norm2.weight

model.diffusion_model.layers.13.q_norm.weight

model.diffusion_model.layers.14.k_norm.weight

model.diffusion_model.layers.14.norm1.weight

model.diffusion_model.layers.14.norm2.weight

model.diffusion_model.layers.14.q_norm.weight

model.diffusion_model.layers.15.k_norm.weight

model.diffusion_model.layers.15.norm1.weight

model.diffusion_model.layers.15.norm2.weight

model.diffusion_model.layers.15.q_norm.weight

model.diffusion_model.layers.16.k_norm.weight

model.diffusion_model.layers.16.norm1.weight

model.diffusion_model.layers.16.norm2.weight

model.diffusion_model.layers.16.q_norm.weight

model.diffusion_model.layers.17.k_norm.weight

model.diffusion_model.layers.17.norm1.weight

model.diffusion_model.layers.17.norm2.weight

model.diffusion_model.layers.17.q_norm.weight

model.diffusion_model.layers.18.k_norm.weight

model.diffusion_model.layers.18.norm1.weight

model.diffusion_model.layers.18.norm2.weight

model.diffusion_model.layers.18.q_norm.weight

model.diffusion_model.layers.19.k_norm.weight

model.diffusion_model.layers.19.norm1.weight

model.diffusion_model.layers.19.norm2.weight

model.diffusion_model.layers.19.q_norm.weight

model.diffusion_model.layers.2.k_norm.weight

model.diffusion_model.layers.2.norm1.weight

model.diffusion_model.layers.2.norm2.weight

model.diffusion_model.layers.2.q_norm.weight

model.diffusion_model.layers.20.k_norm.weight

model.diffusion_model.layers.20.norm1.weight

model.diffusion_model.layers.20.norm2.weight

model.diffusion_model.layers.20.q_norm.weight

model.diffusion_model.layers.21.k_norm.weight

model.diffusion_model.layers.21.norm1.weight

model.diffusion_model.layers.21.norm2.weight

model.diffusion_model.layers.21.q_norm.weight

model.diffusion_model.layers.22.k_norm.weight

model.diffusion_model.layers.22.norm1.weight

model.diffusion_model.layers.22.norm2.weight

model.diffusion_model.layers.22.q_norm.weight

model.diffusion_model.layers.23.k_norm.weight

model.diffusion_model.layers.23.norm1.weight

model.diffusion_model.layers.23.norm2.weight

model.diffusion_model.layers.23.q_norm.weight

model.diffusion_model.layers.24.k_norm.weight

model.diffusion_model.layers.24.norm1.weight

model.diffusion_model.layers.24.norm2.weight

model.diffusion_model.layers.24.q_norm.weight

model.diffusion_model.layers.25.k_norm.weight

model.diffusion_model.layers.25.norm1.weight

model.diffusion_model.layers.25.norm2.weight

model.diffusion_model.layers.25.q_norm.weight

model.diffusion_model.layers.26.k_norm.weight

model.diffusion_model.layers.26.norm1.weight

model.diffusion_model.layers.26.norm2.weight

model.diffusion_model.layers.26.q_norm.weight

model.diffusion_model.layers.27.k_norm.weight

model.diffusion_model.layers.27.norm1.weight

model.diffusion_model.layers.27.norm2.weight

model.diffusion_model.layers.27.q_norm.weight

model.diffusion_model.layers.28.k_norm.weight

model.diffusion_model.layers.28.norm1.weight

model.diffusion_model.layers.28.norm2.weight

model.diffusion_model.layers.28.q_norm.weight

model.diffusion_model.layers.29.k_norm.weight

model.diffusion_model.layers.29.norm1.weight

model.diffusion_model.layers.29.norm2.weight

model.diffusion_model.layers.29.q_norm.weight

model.diffusion_model.layers.3.k_norm.weight

model.diffusion_model.layers.3.norm1.weight

model.diffusion_model.layers.3.norm2.weight

model.diffusion_model.layers.3.q_norm.weight

model.diffusion_model.layers.30.k_norm.weight

model.diffusion_model.layers.30.norm1.weight

model.diffusion_model.layers.30.norm2.weight

model.diffusion_model.layers.30.q_norm.weight

model.diffusion_model.layers.31.k_norm.weight

model.diffusion_model.layers.31.norm1.weight

model.diffusion_model.layers.31.norm2.weight

model.diffusion_model.layers.31.q_norm.weight

model.diffusion_model.layers.32.k_norm.weight

model.diffusion_model.layers.32.norm1.weight

model.diffusion_model.layers.32.norm2.weight

model.diffusion_model.layers.32.q_norm.weight

model.diffusion_model.layers.33.k_norm.weight

model.diffusion_model.layers.33.norm1.weight

model.diffusion_model.layers.33.norm2.weight

model.diffusion_model.layers.33.q_norm.weight

model.diffusion_model.layers.34.k_norm.weight

model.diffusion_model.layers.34.norm1.weight

model.diffusion_model.layers.34.norm2.weight

model.diffusion_model.layers.34.q_norm.weight

model.diffusion_model.layers.35.k_norm.weight

model.diffusion_model.layers.35.norm1.weight

model.diffusion_model.layers.35.norm2.weight

model.diffusion_model.layers.35.q_norm.weight

model.diffusion_model.layers.36.k_norm.weight

model.diffusion_model.layers.36.norm1.weight

model.diffusion_model.layers.36.norm2.weight

model.diffusion_model.layers.36.q_norm.weight

model.diffusion_model.layers.37.k_norm.weight

model.diffusion_model.layers.37.norm1.weight

model.diffusion_model.layers.37.norm2.weight

model.diffusion_model.layers.37.q_norm.weight

model.diffusion_model.layers.38.k_norm.weight

model.diffusion_model.layers.38.norm1.weight

model.diffusion_model.layers.38.norm2.weight

model.diffusion_model.layers.38.q_norm.weight

model.diffusion_model.layers.39.k_norm.weight

model.diffusion_model.layers.39.norm1.weight

model.diffusion_model.layers.39.norm2.weight

model.diffusion_model.layers.39.q_norm.weight

model.diffusion_model.layers.4.k_norm.weight

model.diffusion_model.layers.4.norm1.weight

model.diffusion_model.layers.4.norm2.weight

model.diffusion_model.layers.4.q_norm.weight

model.diffusion_model.layers.40.k_norm.weight

model.diffusion_model.layers.40.norm1.weight

model.diffusion_model.layers.40.norm2.weight

model.diffusion_model.layers.40.q_norm.weight

model.diffusion_model.layers.41.k_norm.weight

model.diffusion_model.layers.41.norm1.weight

model.diffusion_model.layers.41.norm2.weight

model.diffusion_model.layers.41.q_norm.weight

... 60 more

Moss Yuni anime UWU 8 step

sampler

dpmpp 2m_SDE\ er_sde

scheduler

SGM uniform\bong_tangent\beta 57

CFG 1-1.1

DEMO comparison

MOSS SFT + 8-step lora..............................................................................MOSS Yuni Anime UwU turbo

Files with SHA256: 9d3463f5e74008a0dc3128b2351f8016eda58ea0715fb8bf9a688a1fe433a860