AdwolfCzar/krea2-nextscene-raw-v3

Krea 2 "next scene" reference LoRA — v3 (trained on Raw, stock ComfyUI contract)

Given reference image A + prompt, generates the next scene B: same characters, outfit and style, with a new moment, pose or camera.

Use (stock ComfyUI ≥ c9602625; tested on f856877e / v0.39; no custom nodes)

1. Load the model with UNETLoader krea2_raw_fp8_scaled, then LoraLoaderModelOnly krea2_turbo_lora_rank_64_bf16 (1.0). Together these are Turbo.
2. Apply this LoRA with a second LoraLoaderModelOnly (strength 1.0; try 0.7 for freer prompts).
3. Encode with TextEncodeQwenImageEditPlus(clip, prompt, vae, image1=A).
4. Pass the conditioning through FluxKontextMultiReferenceLatentMethod(index_timestep_zero). It is required: without it Krea 2 silently ignores the reference latent.
5. Sample with euler, 8 steps, CFG 1. Output size about 1 MP at A's aspect ratio.

Files

• checkpoints/raw512_r32/stepN: 512 run, 1.5 epochs, r32. s1220 is the 512 pick.
• checkpoints/finish1024/stepN: a short 1024 finish from 512 s1220 (240 steps, LR 5e-5).
• checkpoints/raw512_long3ep/stepN: the same recipe trained for 3 epochs (length test).
• evidence/: eval grids (Turbo 8 steps, seed 76) and configs.
  • Grid columns: native_vl (plain node, no method node), native (index_timestep_zero, no LoRA), A1000 (old Raw-trained LoRA), our checkpoints, strength 0.7, shuffled reference.
  • Grid rows: train pairs with their caption and an altered caption, held-out pairs, short prompts, internet images.

Recipe

• Code: github.com/adbrasi/diffusion-pipe-easycontrol, branch krea2-raw-v3, configs examples/krea2_raw_v3/. Notes: docs/KREA2_RAW_RUN_LOG.md on branch claude/elegant-ptolemy-8kywqy.
• Data: AdwolfCzar/ofificial_next_scene_v11 (natural captions), 1,629 train pairs and 12 held-out.
• Training setup:
  • Base: plain Krea 2 Raw fp8_scaled with its scale. No Turbo LoRA or ostris adapter merged.
  • Reference: frame 1 at t=0 (index_timestep_zero), with Qwen-Image-Edit-Plus VL grounding.
  • Timesteps: krea2_apex law.
  • Real AR buckets, including 16:9.
  • LoRA r32, AdamW8bitKahan, LR 1e-4 cosine, micro 1 × accum 2, 512 px.

Known limits

• The training pairs are consecutive frames: 88% of captions say "the same background". The LoRA keeps the location and is weak at moving to a completely new place.
• Pose, action and camera changes work. Lower strength (0.6–0.75) gives the prompt more room.