k2-proximacena-grounded-v2-full

Next-scene LoRA — a reference-conditioned adapter that teaches an image model to generate the next scene: given a reference image, it keeps character identity, style, palette and world details while the prompt describes the new framing, pose and action.

Training run started 2026-08-04. Checkpoints are uploaded live to checkpoints/step-NNNNN/ as training progresses; per-checkpoint sample grids land in samples/stepN/.

Method

• Pipeline: krea2_omini_grounded from adbrasi/diffusion-pipe-easycontrol, branch ic-lora, commit 8562ac6.
• Technique: sequence [text | noisy target | clean reference]; reference RoPE width-shift, reference modulated at t=0 per-token; condition-only routed LoRA rank 64 on the 28 transformer blocks (delta applied only to the reference rows — zero drift on target rows) plus a global LoRA on the text-fusion stack; visual grounding of the caption against the reference via the VL text encoder (grounding images at 384² max).
• Captions: pure detailed descriptions of the target image only — no instructions, no trigger words, no style words. What the caption omits must be inherited from the reference.
• Optimizer AdamW8bitKahan, lr 1e-4, batch 1 (no gradient accumulation), 1024px area with AR buckets, caption dropout 0.1, bf16 (base weights fp8).
• 15,000 steps planned (~1 epoch of 14,965 samples/epoch), checkpoint + samples every 250 steps.

Dataset

AdwolfCzar/proxima_cena_grounded_original_dataset — 12,455 control/target pairs (the ds4_contexto_curado subset, 1,255 manually curated pairs, is trained with 3x repeats).

Files

• training_config.toml — the config as used (base-model identifier redacted; documentation copy, not runnable).
• dataset_config.toml — dataset config as used.
• sampling_config.json — the exact sampling prompts/references/sizes used every 250 steps (turbo variant, 8 steps, seed 76).
• checkpoints/step-NNNNN/adapter_model.safetensors — LoRA adapters (512 tensors).
• samples/stepN/ — fixed-prompt samples per checkpoint.
• smoke/step12 — 12-step smoke-test adapter (pipeline validation only, ignore).

Inference

Use the runner or the ComfyUI node from the code repo (CtxRush - Krea 2 Omini-Grounded), LoRA applied at runtime in bf16 (never merged into fp8 weights), reference fed as IMAGE+VAE.