Clean Forcing OmniDreams corrector (v2-v3 val-peak)

Trained Clean Forcing drift-corrector LoRA for the OmniDreams single-view integration (deploy at alpha*(t) x 0.25 via OmnidreamsRunnerConfig.drift_corrector).

• lora_v2_v3_valpeak.pt — md5 84beb8c014346833ce182310373b5d32 (88,339,905 bytes)
• demo_sbs_scen7.mp4 — held-out scene, base LEFT vs corrected RIGHT (real-time)
• demo_scen7_4x.gif — the same side-by-side at 4x speed

Style-skin LoRA v6 (4 skins: arcade, comic, cyberpunk, pixel-art)

v6 multi-style game-skin LoRA for the omnidreams live-edit stack — one prompt-switchable adapter covering arcade_racer, comic_ink, cyberpunk_neon (headline: full-hold neon-night conversion), and pixel_art (honest-partial style identity). Trained on teacher-regenerated targets (28 pairs; cyberpunk/pixel_art corpora from layout-preserving JoyAI re-audition). Arcade regression vs v5: none.

Serve with the v5 corrector stack (release style-skin-v5-stack): STYLE_CORRECTOR=lora_style_corrector_v5_valpeak.pt, CORRECTOR_GAIN=0.15, GATE_ALPHA_JSON=gate_style_v5.json, unsharp post. Evidence: NVIDIA#458, live-game integration NVIDIA#494.

Style-skin v5 stack (teacher-regen LoRA + re-paired corrector, final serving config)

Final Issue-1 serving configuration for the omnidreams game-skin edit:

• lora_style_v5_step1600.pt — r64 multi-style LoRA trained on teacher-regenerated (SDEdit sigma 0.65 through the 35-step bidirectional teacher) styled targets. Kills the v3 +10..+14 style melt; ~2x usable style-hold depth with a flat post-swap divergence profile.
• lora_style_corrector_v5_valpeak.pt — rank-16 clean-forcing drift corrector re-paired against v5 (val dag-R^2 +0.368).
• gate_style_v5.json — measured style-specific alpha gate profile (t803 = 0.995).

Serve with: EDIT_LORA= STYLE_CORRECTOR= CORRECTOR_GAIN=0.15 GATE_ALPHA_JSON=, plus unsharp post (1.8/-0.8, sigma 2). Known residual: faint sky banding from the corrector (RMS 0.26 vs 0.12 bare). Alt gain 0.20 trades sky cleanliness for deep-window sharpness. Evidence: PR NVIDIA#458.

Style-skin LoRA v3 (multi-style, maintenance episodes)

r64 edit-timestamped style LoRA for omnidreams live edits (PR NVIDIA#458). Trained on JoyAI restyle pairs (arcade_racer, comic_ink, lowpoly, toy_world full-clip; heavy styles early-window), 1600 steps with 30% style-maintenance episodes. Pre-swap bit-exact; loads through TextEditLoRA (PR NVIDIA#431). Pair with the style-drift corrector release for long holds. md5: a6318aded48d68232902366846d1f0a7

Style-drift corrector v1 (rank-16, gain 0.25)

Clean-forcing corrector for long style-window holds (PR NVIDIA#458): maps drifted styled latents back to the model's own early-window styled manifold. Rank-16 self-attn LoRA in the drift-corrector deploy format (PR NVIDIA#398 hook; run UNFUSED with TextEditLoRA attached first, gain 0.25). Val dag-R^2 +0.42; 20-chunk arcade hold keeps road/lanes/vehicles where the uncorrected hold collapses by +18 chunks. md5: 31ab8776d3a22c5d5841677d02bbad04

Omnidreams text-edit guidance LoRA (r64, step 1600)

Guidance self-distillation checkpoint for NVIDIA#431: plain prompt swaps respond at 0.854 of two-branch guided strength at zero inference cost (deploy via text_edit_lora_path / omnidreams/_edit_lora.py). Trained per integrations/omnidreams/guidance_distill/PLAN.md (r64, 1600 steps, s=3 teacher). md5 1fc7b33ea037f745eeb8d5a1bee59371.

Clean Forcing HY-WorldPlay corrector v2c2 LoRA (shipped)

Supersedes the v2 asset for motion jobs (owner-verified): content-diversified continue-train (200 steps, POOL_WEIGHTS 10% person/scene-diverse pairs). Bridge drift cut −58% (v2: −48%); walking-person scene fixed (drift −47%, best seam profile, owner eyeball pass). Deploy identically: drift_corrector=, alpha*(t) x 0.5 dial for commanded motion, corrector OFF for static. md5 1de6023c92b9d3a967d4bfb694eda235. v2 asset retained under tag clean-forcing-hy-v2 for reproducibility.

Clean Forcing HY-WorldPlay corrector v2 LoRA

v2 corrector LoRA for the HY-WorldPlay integration (PR NVIDIA#396). Deploy: set drift_corrector to this file; the runner applies alpha*(t) x gain per step (shipped gain 0.5) with content-keyed selection (static trajectories run the base). Trained zero-real via the Clean Forcing recipe in integrations/hy_worldplay/drift_correction/.