ACCC1380/shuiwu22-moss-image-lora

shuiwu22 MOSS-Image LoRA — fresh 120-epoch training

Latest final adapter: shuiwu22-moss-image-120epoch.safetensors. Trigger: shuiwu22.

This run starts with freshly initialized LoRA weights on the original, nonquantized MOSS-Image-Base. It does not continue the earlier 600-step adapter. Earlier shuiwu22-moss-image-lora files remain as separate alternatives.

Training

76 unique images, 120 full epochs, 9,120 sample presentations. Physical batch size 1 and gradient accumulation 4 give effective batch size 4, 19 optimizer updates per epoch, 2,280 total updates. The trainer supports only physical batch 1 because captions and image buckets have variable lengths.

1024-area aspect-ratio buckets, no upscaling; rank/alpha 32 on attention and feed-forward projections of all 42 blocks; dropout .05. AdamW, peak LR 5e-5, weight decay .01, cosine decay, 95 warmup updates, gradient clipping 1, seed 42. Released original bf16 transformer, native text encoder dtypes, fp32 VAE; fp32 adapter and optimizer state. No quantization. Text encoder and VAE frozen. Flow-matching velocity MSE, logit-normal sampling (-1.09,.95).

Existing dual-WD captions are reused unchanged: WD EVA02-Large v3 and ViT-Large v3, one pass each, normalized case-insensitive tag union without duplicates. General thresholds .5296 and .35; character threshold .85; rating labels excluded. No new tagging was performed.

Checkpoints and use

Final adapter is epoch 120; files ending in 000010 through 000110 preserve every tenth epoch. These alternatives allow comparing adaptation strength. The 120-epoch target follows the user's explicit instruction; more epochs do not establish better visual quality.

Use a current MOSS-compatible ComfyUI and LoraLoaderModelOnly. Trainer guidance reports Base-trained adapters work with both Base and SFT. Initial inference settings: strength .5-.7, 50 steps, CFG 3, Euler, shift 5, negative-branch block 22 skipping. These starting settings have not been visually tested for this adapter.

Validation

See validation.json for exact epochs, steps, finite tensor checks and SHA256 hashes. Only adapter weights and reproducibility records are uploaded. Images, captions, private filenames and image/text caches are excluded. The assistant did not view training images, captions or previews. No visual fidelity, generalization or overfitting assessment was performed. Numerical stability does not prove sufficient fitting or rule out overfitting.

Training implementation: PineCookie/LoRA-Forge. Original MOSS and WD model provenance is in model-provenance.json; software versions are in environment.json. Upstream model licenses do not establish rights to training data; no new dataset or adapter license is asserted here.