4K image editing in 4 steps, by swapping the VAE

What it does BluePixel is an instruction-based image editor: you give it a photo and a prompt like “change the gray sweater to burgundy red,” and it returns an edited image at up to native 4096×4096 — no cascaded super-resolution stage. It’s built on FLUX.2 Klein (a 9B DiT with Qwen3 as text encoder) and ships as a 4-step distilled checkpoint, BluePixel-Turbo, distilled from a 50-step teacher. The repo is inference code only; training is described but not included.

The interesting bit The trick that makes 4K tractable is the VAE swap: replacing Klein’s f=8 VAE with DCAE-F64 (f=64, 1×1 patchify) shrinks the 4K latent from ~65,536 tokens to ~4,096 — the same token count as 1K editing. The DiT never sees a bigger problem. Distillation uses source-anchored target projection (SATP), which anchors non-edited regions to the source latent so backgrounds stay put across just 4 steps.

Key highlights

• Native 4096×4096 output; non-square images keep roughly their aspect ratio within a requested pixel budget (target_pixels).
• 4-step inference, ~8.7 seconds per 4K edit on a Kunlunxin P800 XPU, ~4× faster than the 50-step teacher.
• 4K adaptation took ~200K samples on 32 Kunlunxin XPUs in about 4 days, via progressive compression migration (f=8 → 32 → 64).
• Continual learning via ACE-LoRA + OPD absorbs long-tail editing skills without catastrophic forgetting; RL alignment uses PickScore and a custom BluePixReward.
• Batch runner (vins_infer.py) has resume-by-default, file-lock multi-process support, and dry-run/check modes — clearly built for real evaluation runs.

Caveats

• The evaluation table in the README is empty — no benchmark scores are actually published, so quality claims rest on the demo and paper.
• Requires CUDA with BF16, targets the Diffusers 0.36 API, and the README itself notes GPU inference must be validated in your deployment environment.
• Training code isn’t released; this is inference-only, and the disclaimer asks for security review before product use.

Verdict Worth a look if you edit high-resolution images and want 4K output without a super-resolution pipeline — the token-count trick is genuinely elegant. Skip it if you need training code or published benchmark numbers to justify a pick.