slopshop/erase-v1

erase_v1 — object removal for FLUX.2 [klein] 4B turbo

A LoRA for FLUX.2 [klein] 4B (the distilled "turbo" variant: 4 steps, no CFG), and the precomputed prompt embedding it is used with. Given a photo and a selection, it replaces the selection by the surrounding background, object shadows and reflections included.

It powers the generative fill of SlopShop, an open-source image editor, which runs it locally through ONNX Runtime.

Files

File	Content
erase_v1_diffusers.safetensors	the LoRA, diffusers keys (transformer.transformer_blocks.N.attn.to_q.lora_A.weight…), rank 32, alpha 32, scale 1.0, 100 layers
prompt_embeds.safetensors	prompt_embeds bf16 [1, 512, 7680] and text_ids int64 [1, 512, 4]: the prompt below, encoded once, so that no text encoder is needed

Prompt (for information; do not re-encode it): Remove the object under the white area of the mask (second image), including its shadow and reflection, and fill that area naturally with the surrounding background.

Use

• Working size: at most 1 Mpx, both sides multiples of 16.

• Reference images:

  1. the photo, unchanged;
  2. the mask as an RGB image (white to remove, black to keep).

  Both are VAE-encoded (the distribution's mean), 2×2-patchified and normalized by the VAE's batch-norm statistics.

• Joint sequence: [target, photo, mask]. Their position ids are (T, y, x, 0), with T = 0, 10 and 20 respectively. The text ids are (0, 0, 0, l).

• Steps: 4 Euler steps on FLUX.2's resolution-shifted sigmas, one pass per step, no guidance.

• Composite: strictly the result inside the selection and the input outside it.

• Selection: grow it by about 2 % of the larger side, or use its convex hull; a selection tight on the object can leave a faint remnant.

Provenance and license

Apache-2.0.

• Distilled from FLUX.2 [klein] base 4B and fal's object-remove LoRA, both Apache-2.0.
• Trained on images of the CORNE set (Apache-2.0).