MicksHF/fusal-refchar-v17-stage2-3

fusal-refchar-v17 Stage2 and Stage3

This repository contains the Stage2 and Stage3 DoRA checkpoints for the fusal-refchar-v17 reference-character edit training run.

Recommended checkpoint

Use the final Stage3 checkpoint for high-resolution testing:

stage3/refchar_v20_stage3_dora64_960_bf16_96gb.safetensors

Stage2 is included so the training trajectory can be reproduced and compared:

stage2/refchar_v20_stage2_dora64_768_bf16_96gb.safetensors

Contents

• stage2/: Stage2 DoRA checkpoints, including final and retained late checkpoints.
• stage3/: Stage3 DoRA checkpoints, including final and retained late checkpoints.
• configs/: AI Toolkit configs used for the run and smoke test.
• logs/: training logs for Stage2 and Stage3.
• samples/: generated sample images and Stage3 trajectory comparison images.
• docs/: training notes, file manifest, and SHA256 checksums.

Training summary

Base model:

• Qwen/Qwen-Image-Edit-2511
• Local base used during training included the prior Stage1 v17 merge.

Dataset:

• YuHuison/fusal-refchar-v20
• 10,000 target/reference/caption pairs converted to AI Toolkit format.

Network:

• AI Toolkit DoRA
• rank 64, alpha 64
• attention projections only: to_q, to_k, to_v, to_out, add_q_proj, add_k_proj, add_v_proj, to_add_out
• bf16 checkpoint dtype

Stage2:

• resolution: 768
• steps: 10000
• optimizer: adamw8bit
• LR: 1e-4
• noise scheduler: flowmatch
• timestep type: weighted
• caption dropout: 0.05
• samples: 768x768, every 500 steps

Stage3:

• warm-started from the final Stage2 DoRA checkpoint
• resolution: 960
• samples: 1280x720, every 1000 steps
• steps: 10000
• optimizer: adamw8bit
• LR: 3e-5
• noise scheduler: flowmatch
• timestep type: weighted
• caption dropout: 0.05

For 16:9 images, AI Toolkit resolution 960 maps to the 1280x720 bucket.

Notes

The training run did not use negative prompts. For this diffusion training setup, the model learns from the reference image and positive edit caption condition. Negative prompts are mainly an inference-time steering tool, not a required training field.

See docs/TRAINING_NOTES.md for the trajectory notes and operational details.