Downloads · 30 days
0
Zaytron40k/Qwen-Image-Edit-2511-SeamRemove-LoRA
Qwen-Image-Edit-2511-SeamRemove-LoRA is a image-to-image model from Zaytron40k. Use it when you need one image transformed into another. It is set up for diffsynth-studio.
A corrective edit LoRA that heals rectangular editing seams: feed an image where a cropped region was pasted back with a hard seam / color drift / warp / offset (e.g. the aftermath of an inpaint), and it outputs the c…
Downloads · 30 days
0
Access
Public
Updated Jun 19, 2026
Repo size
2.8 GB
Likes
1
Public
Click a slice to open those files.
.safetensors2.8 GB · 100%
From the Hugging Face model README
A corrective edit LoRA that heals rectangular editing seams: feed an image where a cropped region was pasted back with a hard seam / color drift / warp / offset (e.g. the aftermath of an inpaint), and it outputs the clean, seamless scene. Trained with DiffSynth-Studio on 1685 synthetic pairs (337 clean scenes x 5 corruptions: random/face/hand placements).
Trigger / instruction: s3amfix
s3amfix, remove the rectangular editing seam and blend the region seamlessly, matching color, contrast and alignment
(optionally append a scene description). Trained dual-caption 50/50 (instruction-only vs instruction+scene).
img = pipe("s3amfix, remove the rectangular editing seam and blend the region seamlessly, matching color, contrast and alignment",
edit_image=[seamed_image], seed=0, num_inference_steps=40, zero_cond_t=True) # zero_cond_t REQUIRED
| base | Qwen/Qwen-Image-Edit-2511 (DiT only) |
| rank / lr | 32 / 7e-5 |
| epochs x steps | 6 x 1685 (1685 pairs, repeat 1) |
| resolution | input & target identical-dim, max_pixels 1763584 (~1328^2) |
| precision | bf16 + gradient checkpointing |
| special | --zero_cond_t (2511-specific, also required at inference) |

| epoch | step | EMA loss min |
|---|---|---|
| 0 | 763 | 0.0166 |
| 1 | 3006 | 0.0179 |
| 2 | 4225 | 0.0170 |
| 3 | 5633 | 0.0152 |
| 4 | 7485 | 0.0156 |
| 5 | 8637 | 0.0145 <- global min |
Loss is flat/noisy (normal) - pick the checkpoint by visual eval on WaveSpeed, not by loss.