Implement the Lightricks two-stage recipe (diffusers PR #13217) for the
LTX-2.x Dev family: Stage 1 at half-res with full four-way guidance,
2x latent upsample, Stage 2 with distilled LoRA + scheduler swap + identity
guidance on STAGE_2_DISTILLED_SIGMA_VALUES.
Extends to both LTX-2.0 and LTX-2.3 Dev via per-family distilled-LoRA
repos carried on the caps; Distilled variants take the same flow minus
the LoRA swap. Auto-couples Refine with a fixed 2x upsample on any Dev
variant with a known LoRA when the user enables Refine without Upsample.
- caps: is_ltx_2_3, use_cross_timestep, default_dynamic_shift,
stage2_dev_lora_repo, supports_canonical_stage2, modality_default_scale,
guidance_rescale_default; LTX-2.x defaults realigned to canonical
cfg=3.0 / steps=30; per-variant STG block and four-way guidance wired
for non-distilled 2.x
- process: canonical Stage 1/Stage 2 helpers, scheduler + opts snapshot
under try/finally, per-family upsampler repo, audio latents threaded
from Stage 1 into Stage 2, use_cross_timestep gated per caps
- overrides: skip the redundant unsharded LTX-2.3 connectors blob and
share LTX2TextConnectors weights across 2.3 variants when te_shared_t5
- load: Gemma3 shared-TE path for LTX-2.3; gate use_dynamic_shifting=False
override to 0.9.x only so LTX-2.x stays on its canonical token-count
dynamic shift
Move cache tracking from ltx_util into video_load where shared.sd_model
lives, and invalidate the name-based hit when the cached class no longer
matches the current pipeline (e.g. after Unload Models triggers an
auto-reload of the default checkpoint).
- Drop the duplicate module-level loaded_model cache in ltx_util
- Add a pipe-class isinstance check around the cache hit in video_load
Rework the LTX Video tab so one UI handles every registered variant
(0.9.0 through 2.3, Dev/Distilled/SDNQ-4Bit, T2V/I2V/Condition). Per-
variant behavior is driven from a single capability lookup rather than
substring matching on model names scattered across the backend.
- modules/ltx/ltx_capabilities.py: new module computing family, is_i2v,
distilled, supports_input_media, supports_multi_condition,
supports_image_cond_noise_scale, supports_decode_timestep,
supports_stg, supports_audio, supports_frame_rate_kwarg, and the
default CFG / steps / sampler_shift for a given model name by reading
its registered repo_cls in models_def.
- modules/ltx/ltx_ui.py: capability-gated UI. Selecting a model rewires
accordion visibility, slider interactivity, and defaults via a single
model.change handler. New controls: dedicated image input slot inside
the LTX tab (replaces the disconnected shared init_image for I2V),
condition strength slider, CFG / sampler shift / dynamic shift
sliders that were previously unreachable. Input media accordion
restructured so the image slot is always-visible while the video /
gallery prefix tabs only appear on Condition pipelines.
- modules/ltx/ltx_process.py: route the base pass through
processing.process_images(p) so LTX inherits standard scheduler
wiring, extra_networks activation, VAE handling, and error plumbing
from StableDiffusionProcessingVideo. The multi-pass latent path
(upsample / refine) stays on direct pipeline calls for latent
re-entry. Refine noise control gets family-specific kwargs:
denoise_strength for 0.9.x LTXConditionPipeline, noise_scale for all
2.x pipelines; the prior strength= injection crashed on 2.x and only
affected conditioning intensity on 0.9.x. Add torch_gc between every
stage boundary (base to upsample to refine to vae decode) so the
CUDA allocator cache does not retain the prior pass's allocations
across stages. Remove the TypeError fallback that silently passed
raw latents to save_video when VAE decode returned None on OOM;
those errors now surface cleanly.
- modules/ltx/ltx_util.py: get_conditions grows a family parameter and
builds LTX2VideoCondition (frames, index, strength) for 2.x or
LTXVideoCondition (image, video, frame_index, strength) for 0.9.x.
get_bucket floors to max(32, vae_spatial_compression_ratio) since LTX
pipelines validate divisibility by 32 regardless of family.
- modules/video_models/video_overrides.py: extend the I2V generator
reset to cover LTX2ImageToVideoPipeline and both Condition classes.
Keep the strength= kwarg injection gated to 0.9.x
LTXConditionPipeline only; LTX2ConditionPipeline.__call__ does not
accept it (per-condition strength lives on the LTX2VideoCondition
dataclass instead).
The 2.3 I2V variants (Dev, Distilled, SDNQ-4Bit, Distilled SDNQ-4Bit)
were registered against LTX2Pipeline, whose __call__ does not accept an
image kwarg; init images were silently dropped. Route them through
LTX2ImageToVideoPipeline instead.
Also register the four 2.3 Condition variants under LTX2ConditionPipeline
so the multi-condition (image/video/gallery prefix) generation path is
reachable on LTX 2.3.
Replace "init image not set" / "last image not set" with actionable
messages that tell the user what to do. Matches the phrasing used by
the Caption tab for the same failure class.
CivitAI removed enum validation from /api/v1/models?baseModels=X (now
returns 200 with empty items on unknown values), which left the
base-model filter dropdown empty. Fix in two parts.
1. Point the base_models discovery probe at /images instead of /models.
/images still returns a ZodError with the full enum list (82 names
at time of writing). The existing parser handles the shape unchanged.
2. Also fetch civitai/civitai's base-model.constants.ts from GitHub and
merge per-name metadata (group, ecosystem, engine, family, hidden)
into a new base_models_info field on the response. The live probe
remains authoritative for which names exist; GitHub provides the
metadata and doubles as a fallback name list if the probe comes back
empty.
Backward compatible: the existing base_models list keeps the same
shape. Clients that want grouped dropdowns, hidden-flag filtering, or
image/video separation can consume base_models_info instead.
The GitHub fetch has a separate 6h cache since these constants change
much less often than the probes it sits alongside.
Add three adapter families to the z-image native loader and chain them
with the existing lora path through load_safetensors. Mixed-family
files (for example gta6_amateur_photography_zimagebase_v2.safetensors,
which carries lora and lokr groups in the same file) now load
completely instead of having one family silently dropped.
Shared helpers in pipelines/z_image/zimage_lora.py parse keys by
suffix list, rename legacy attention.out and attention.wo to
attention.to_out.0, and split fused attention.qkv into to_q/k/v. For
lora the split chunks the up weight along dim 0. For lokr the split
emits three NetworkModuleLokrChunk entries that share the tensors and
slice the kronecker product at apply time.
Fused attention.qkv for loha and oft is skipped with a warning. No
NetworkModuleHadaChunk exists, oft rotations are tied to out_features
and cannot be cleanly split across q/k/v, and no real z-image adapter
in that layout exists today.
load_safetensors for zimage chains try_load_lora, try_load_lokr,
try_load_loha and try_load_oft and merges their module dicts into a
single Network so mixed files load every module.
Add zimage to allow_native so lora_force_diffusers picks between
native and diffusers. Before this, zimage always took the diffusers
path regardless of the setting.
pipelines/z_image/zimage_lora.py reads the safetensors and writes
directly into network_layer_mapping, so Z-Image LoRAs no longer go
through the diffusers PEFT converter that raised KeyError on
state dicts with partial alpha keys.
Key formats handled: ai-toolkit, kohya lora_unet_, bare transformer.
and no-prefix. Pre-refactor fused attention.qkv is split into
to_q/k/v; attention.out and attention.wo are renamed to
attention.to_out.0. Alpha and dora_scale are preserved.
upstream UniPCMultistepScheduler.set_timesteps unconditionally moves self.sigmas to CPU after building them. multistep_uni_p_bh_update / multistep_uni_c_bh_update then constructs a torch.ones(..., device=sample.device) tensor and calls torch.stack([..., self.sigmas[...]]) — crashing at inference step >= 2 whenever the model runs on a non-CPU device (CUDA, ROCm, MPS).
Monkey-patch set_timesteps so that, after the upstream call, self.sigmas is moved back to the requested device. Applied once at import time inside the existing sampler-load try/except block so failures are silent-logged and never break the rest of the sampler registry.
CalamitousFelicitousness
Disty0
vladmandic
Oleksandr Liutyi
awsr
resonantsky