Lazy text-encoder load: build only on conditioning-cache MISS (saves 21-33GB host RAM + load time on cache-HIT runs)

hires-fix: 33f windows + zero resident blocks during the pass (VAE encoder fatal abort at 65f/1920x1088 with 12GB pinned)

fix: _ref_sched unbound on no-reference runs (no-character scripts)

encoder_fp8: strip unused vision tower + absolute-headroom fits check -> GPU encode engages on 24GB cards

conditioning disk cache (encode once per script per machine) + CPU-fp8 encoder warning

README: document memory-trim fix (critical), fp8_scaled_mm, resident_blocks, encoder_fp8, hires-fix, ref scheduling, robustness; full file list + no-cherry-pick warning

per-device capability probe: pre-Ada GPUs auto-fall back to upcast path

fp8_scaled_mm: clear sm_89+ gate (RTX 30-series lacks fp8 cores)

harden pipeline dtype derivation against fp8 params (2 sites)

harden pipeline dtype derivation against fp8 params

hires-fix second pass: per-shot latent re-encode + DMD tail re-denoise at target res (real detail synthesis; 65f windows, 3090-safe)

fix: scaled-mm swap ran on META model (0 linears matched); branch on weight dtype at call time

guard: fp8-compute toggles require bf16 source file (fp8 files load all-fp8 with cast skipped -> randn crash)

encoder_fp8: halve the Gemma wrapper footprint (fp8 storage, upcast at encode; enables GPU encode pass on 32GB)

torch._scaled_mm fp8 Linear forward (measured x2.8 kernel / x1.5 e2e vs bf16 on RTX 5090)

QuantizationPolicy.fp8_scaled_mm_torch (torch-native, no tensorrt_llm)

fp8_scaled_mm mode (native fp8 matmul, no upcast tax) + resident_blocks partial offload

FIX: memory bank trim was a no-op when max_size<=num_fix_frames (tail[-0:] returns whole list) -> bank grew unbounded

RefPicker: re-entry ref scheduling (3+ shot absence) + script-carried ref pinning via refs map

Generate: schedule-aware ref dedup, cap 6 scheduled entries

TextEncode: hot-swap encoder to GPU for the encode pass (fits-check + OOM fallback)

README: document fp8 gemma scale-key layout fix

fp8 gemma: accept .weight_scale layout + zero-match warning

per-character reference scheduling: RefPicker emits ref_shots (first-appearance shot per character), Generate injects each ref at its character's entry shot with the long trim - multi-character single-run videos get every character properly seeded

load order: DiT (quantize/fuse peak) first, VAEs, text encoder LAST + gc between stages - encoder-first stacked 24GB under the fuse peak and caused an mmap access violation at VAE load on a 96GB box

references seed shot 1 only: from shot 2 the bank carries identity+staging; re-injecting a wide full-scene ref every shot made the subject relocate between shots

RefPicker auto mode: emit ONE ref per character named in prose (batch output) - solo scene 1 ref, two-character 2 refs, none 0; dialogue mentions still excluded; single picker now covers every queue case with zero mode switching

RefPicker prompt scan: parse JSON script payloads before scanning - on raw JSON the plain-quote stripper deleted the entire prompt body, so auto mode never matched a character on TXT-wrapped JSON blocks

RefPicker IS_CHANGED: guard None from unresolved wired inputs (per-item WARNING spam + lost caching)

no-offload pre-flight: refuse cannot-fit DiT with a catchable error (cudaMallocAsync otherwise hard-aborts the whole process), warn on tight activation budgets

reference_zoom: over-zoom refs before conditioning - a ref at exactly render size is pixel-continuable and the model opens the shot ON it (long literal ghost no trim covers); 1.2 default breaks the continuation shortcut while keeping identity + wide composition

head_trim_first_shot: shot 1 conditions on refs alone (empty bank) and opens ON the reference - give it its own longer trim (auto = 2x head_trim, min 24)

LoRA fusion: map kohya-FLATTENED keys (lora_unet_*, dots->underscores) onto model paths via exact flat-key lookup - such LoRAs previously matched zero keys

RefPicker: add '(none / disabled)' character_pick option - disable a spare picker without rewiring

LoRA hardening: kohya lora_down/up support + alpha scaling, fuse telemetry + zero-match warning, refuse comfy_quant checkpoints with clear error; lora_file dropdown

RefPicker: character dropdown, no silent scan-fallback, cache fix; head-trim applied to all outputs; lora_file dropdown on ModelLoader; gemma_path guard

Upload 11 files

initial commit