Krea2 Image Editing LoRA Open-Sourced
molbal · reddit · 2026-07-14
A developer ran out of GPU budget while training a Krea2 image editing LoRA and decided to open-source the project, datasets, and training details, inviting others to carry the baton.
The post shares several technical insights:
- Adopts the indextimestepzero approach from Ostris's ai-toolkit, allowing reference image tokens to participate in modulation at timestep=0
- Places reference tokens on the right side of the target image's RoPE 2D grid to avoid grid artifacts caused by a standalone axis
- Requires a fixed reference label format for the text encoder: <referenceN><|visionstart|><|imagepad|><|visionend|></referenceN>
- Necessitates changing the VLM pixel budget during training from 384x384 to 512x512
Additionally, two Hugging Face datasets were released: one containing about 20,000 real-world semantic editing pairs to teach the model the concept of "editing," and another featuring algorithmically generated identity-preserving/pixel-level preserving data, emphasizing that areas not mentioned in the prompt should be copied exactly from the reference image.
More from Multimodal
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26