Spaces:
Running on Zero
Running on Zero
Commit History
-
Update app.py Running
Verified signature9784675 -
Update app.py
Verified signature5f19892 -
Update app.py
Verified signature947a3b4 -
Update app.py
Verified signature964c9f7
-
Update app.py
Verified signature2e3b82d -
Update app.py
Verified signature26f474d
-
add keyframes, singlepass checkbox, settings import/export
5fab075
-
update transition defaults
1c1fd11 -
add transition lora
3412b17 -
adjust resolution parameters
39a6dc3
-
update resolution defaults
230086f -
fix rounding error
21e59ed -
update resolution handling
0cb921b -
add optional lora support
be3c3af
-
Unified presets with sigma textbox
bd83937
-
Tune identity defaults
Lower OmniNFT default strength and raise latent anchor cache warmup default.4598920
-
Add scene chaining and audio reference prep
Use PromptRelaySmartEncode for relay prompts, add a local audio prep node for reference clipping and normalization, and add an experimental FunPack scene-chain path that bypasses pass 2 for first tests.2e705f0 -
Add audio reference + per-lora audio strength via MultiLoRALoader
Rewrite rgthree Power Lora Loader to phazei MultiLoRALoader in LTX mode so each lora has separate video / audio strength multipliers. Add LTXVReferenceAudio node injection with optional MelBandRoFormer stem separation for voice ID transfer. Exclude KV node from POS/NEG consumer redirect to avoid cycle when KV + audio_ref both enabled. Add rotary_embedding_torch dependency.20a97b5 -
Add mistic better motion + physics v2 loras, sulphur defaults 0.15/0.15, log gen settings
ab94d59 -
Pre-populate workflow cache in parent so fork COW skips re-conversion
1edd609 -
Fix kijai omninft bf16 URL case (repo + filename)
83950c3 -
Expose OmniNFT + sulphur v1 loras, two-regime GPU budget, comment trim
- Move OmniNFT (vasiliyweb converted) and add OmniNFT RL bf16 + sulphur experimental v1 to OPTIONAL_LORAS with user-facing sliders. Default strengths preserve prior behavior (vasiliyweb=0.8, new ones=0.0). - Strip OmniNFT entries from workflow templates during conversion so the sliders are the sole authority on those loras. - Two-regime get_gpu_duration: tight 30+5*work when it fits the 120s/day free ZeroGPU allowance, else fall back to the older proven 45+8*work formula for longer gens. - Targeting accordion default to closed. - Trim over-verbose comments while preserving crash-site debug breadcrumbs.84f250e -
Expose K/V strength as a tunable slider in a new funpack accordion
eeb028f -
Cache K/V zero-prefix tensors to fix VAE NVML asserts; drop kv checkbox info blurb
1fe2a31 -
Use delegating CompressedTimestep subclass when n_ref misaligns with ppf
In pass-2 tile sampler the captured pass-1 ref token count doesn't divide the tile's patches_per_frame. Instead of physically expanding the 24576-wide modulation storage to per-token form (~512MB per call, exhausts VRAM), wrap the original timestep in a same-class subclass that shares the compressed data and prepends n_ref zero rows lazily inside expand_for_computation. Aligned pass-1 case keeps the cheap direct compressed extension.4f5c3f7 -
Fix K/V timestep extension to match vx exactly when n_ref not divisible by ppf
a93addd
-
Extend v_cross_pe in K/V AV compat patch for audio-to-video cross-attention
f594437 -
Patch FunPack KV wrapper for LTX-AV: extend modulation tensors alongside positional embeddings
4088156 -
Add K/V identity conditioning via FunPack wrapper node
1071912 -
Hide input mode UI, revert title to i2v, raise dreamly/singularity defaults
d953a2c -
Add prompt relay (timeline-based prompts) with strict parser and silent fallback
0f224d2 -
Add singularity lora as optional toggle, default off
66762b2 -
Pass all reference images to enhancer; mark MSR modes WIP; reorder input mode radio
cc0ebbb -
Add multi-reference (original) input mode with bundled workflow + normalize lora path separators
d0a3178 -
Route MSR conditioning from IC-LoRA guide direct; match canonical per-guide strength; normalize lora path separators
08b9869 -
Fix MSR conditioning chain and add per-image guide multi
95207ed -
Drop Gemma enhancer, fix MSR node ids (552->740, 556->596)
Gemma 4 26B heretic over-thinks past the GPU duration; sulphur works great for both i2v and MSR per testing, so removing gemma entirely: - delete GEMMA_* constants and GEMMA_MSR_SYSTEM_PROMPT - revert _ensure_enhancer / _start_enhancer_server to single-model - drop enhancer_model dropdown + handler outputs in UI - _enhance_prompt_impl no longer takes model arg MSR injection was failing with 'required workflow nodes missing' because: - node 552 (VAEDecode) is the PASS-1 debug decoder and lives in skip_ids - it never makes it into the API workflow. - node 556 (LTXVSeparateAVLatent) is pass-1 too; the FINAL output flows through node 596 (pass-2 separator) and node 740 (pass-2 decoder, the actual active VAEDecode). Fix: target MSR_NODE_FINAL_SEPARATE=596 and MSR_NODE_VAE_DECODE=740. Also emit a more informative error listing which node ids are missing when the precondition fails.fff6d80 -
Pre-download Gemma enhancer at startup so no model downloads inside GPU fork
d5c31d4 -
Add MSR multi-reference mode and selectable Gemma-4 enhancer for MSR
a9b1948
-
Tune defaults: sulphur 0.3, vbvr 0.5, similarity 0.3, sigmas tuned
8919d44 -
Register sigma change handler last so generate/enhance fn_indexes stay stable
09bf0c4 -
Add refine sigma preset selector (original/tuned/custom) with validation
0535943 -
Remove audio toggle, native AV always on
2ef7b49 -
Expose identity anchor tuning, auto-align cache step to frame count, drop audio gpu-time penalty
85fd13d -
Switch vbvr to archive mirror, add dreamly/synth/plora optional loras
04b50e8
-
Fix get_gpu_duration signature to match generate lora args
74e8179 -
Rename optional lora to sulphur experimental, add vbvr lora
5394928 -
Add optional eros experimental lora, zerogpu budget controls, updated header
cc6cd73 -
Set enhancer GPU duration to 80s
1e5b86a -
Bump enhancer GPU duration to 120s for cold model load
e51c652