The crash fix for the scene-2+ patchify_video failure had been applied in
the shared EasyMultiTrackProject pipeline, changing the hidden first-pass
size of the certified MultiTrack Project and Moxie MultiTrack Project
nodes too. Certification requires those nodes to keep their exact
behavior, so the snap now lives behind overridable hooks whose defaults
are the certified rules:
- _h3_first_pass_size default: max(32, int(editor * scale)) plain
halving (the certified formula). MoxieProjectSlow overrides it to snap
onto the nearest 64px grid, fixing the odd-latent crash on its own
continuation segments only (1080p -> 960x512, 1344x768 -> 704x384,
4K -> 1920x1088 unchanged).
- _h3_needs_latent_upscale default: the certified 'upscale_by <= 1'
skip rule. MoxieProjectSlow overrides it with a true size comparison
(load-bearing for its latent-upscale refine at upscale_by 1.0).
- The certified sampling preview sampler keeps its widget text, default
2 and max 8; min stays 0 only because the engine passes the Slow
node's automatic tile_count 0 through it and ComfyUI validates inputs
against min at execution.
utils/h3_project.py is back to its certified state (the snap helper now
lives in the Slow node's own file). Regression tests pin the certified
first-pass dims and decision rules of both certified project nodes so
any future drift fails the suite loudly.
Multi-segment projects crashed on the first continuation segment with
RuntimeError: shape '[1, 24, 1, 1, 8, 2, 15, 2]' is invalid for input of
size 12240 (patchify_video). The hidden half-resolution first pass
declared its size with plain integer halving, so a 1920x1080 editor
sampled at 960x540 - and a latent encoded from 540px carries an odd
17-row count. Odd latents sample fine, but when the segment tail comes
back as motion-context keyframes on the next segment, the model's
patchify_video (2x2 latent patches) cannot reshape them.
The first pass now snaps onto the nearest 64px grid (h3_even_latent_size)
so every latent in the pipeline keeps even row/column counts:
- 1920x1080 editor samples at 960x512 (was 960x540 -> 17 rows)
- 1344x768 samples at 704x384 (672 had odd latent columns)
- 3840x2160 is unchanged at 1920x1088 - the size the VAE already rounds
up to, so the 4K-editor trick behaves identically
Outputs and second-pass dimensions are unchanged; only continuation
segments that previously crashed now run. Also updates the stale
single-task expansion test, which still expected pre-half-res-trick
conditioning dimensions (editor size instead of the sampled size).
Tiling stops being a baked setting and becomes a widget: the new Boolean
tiling switch next to upscale_model (default on) turns the memory-adaptive
spatial tiles of the 2-step refine pass off, so the pass can be sampled in
one piece on machines that prefer the pre-tiling behaviour.
- OFF injects {enabled_tiling: "false", tile_count: 0}; the engine then
builds the plain SamplerCustomAdvanced second pass instead of the tiled
preview sampler
- ON keeps tile_count 0 (automatic): the engine picks the tile count from
the available accelerator memory and skips tiling when the full
evaluation fits, keeping 1920x1080 within reach of 12 GB and AMD/Intel
GPUs
- workflows saved before the switch existed keep tiling on
- tests: schema slot order + defaults, switch ON/OFF passthrough, legacy
fallback; docs: CHANGELOG + multitrack-schema reference
Bake enabled_tiling with tile_count 0 (automatic) into the Slow node's
second pass: the engine picks the tile count from the available
accelerator memory and bypasses tiling entirely when the full evaluation
fits. This brings 1920x1080 output within reach of 12 GB cards and
AMD/Intel GPUs - the tiling machinery only uses device-agnostic ComfyUI
memory helpers.
- _h3_tiling_config now accepts tile_count 0 (automatic) alongside 2-8
- easy h3SamplingPreviewSampler schema allows tile_count min 0
- tests: tiling config parsing, baked Slow passthrough, second-pass
graph receives tile_count 0, sampler schema range
- docs: CHANGELOG + multitrack-schema reference
The Slow node previously baked the upscaler checkpoint filename
(minimax_h3_latent_upscaler_3d_fp16.safetensors) into its settings with
no place on the node to choose it. The checkpoint is now a proper
upscale_model combo widget placed with the model components (after
audio_vae):
- options are None plus every file under ComfyUI/models/latent_upscale_models/
(the bundled default is always kept selectable so the combo default is
valid even before the checkpoint is downloaded)
- picking None (or an empty string, e.g. from a converted primitive
socket) raises a clear error instead of silently falling back to the
decode/RTX-resize/re-encode path this node exists to avoid
- workflows saved before the widget existed fall back to the bundled
checkpoint, so they run unchanged
- engine untouched: the value flows through the existing
kwargs.get('upscale_model') read; baked settings simply stop
overriding it (base BAKED_PROJECT_SETTINGS copy pops the key)
- MoxieProject gains a _component_inputs() schema hook so subclasses can
slot inputs in with the model components; base node schema unchanged
- tests: schema spot/ordering, folder listing without duplicates, widget
passthrough, pre-widget fallback to the bundled checkpoint, None/empty
rejection; CHANGELOG + multitrack-schema.md updated
New project node variant (MoxieProjectSlow, display name 'Moxie
MultiTrack Project Slow') that replaces the RTX VSR pixel restore with
a latent upscaler plus a short refine pass:
- identical UX to MoxieProject: end users only set steps and seed (plus
the shared context_frames overlap); schema mirrors the base node via
a shared _project_inputs() helper
- first pass stays at the hidden half resolution (engine gains a
_h3_first_pass_resolution_scale class hook, overriding the previous
sampling_mode == 'single' hardcode; MoxieProjectSlow returns 0.5
under dual mode)
- the half-res latent is upscaled to the final (editor) size by the H3
latent upscaler (minimax_h3_latent_upscaler_3d_fp16.safetensors,
models/latent_upscale_models) and refined with a hidden 2-step euler
pass at 0.33 denoise; the schedule replicates KSampler.set_steps
(int(steps/denoise) total steps, last steps+1 sigmas) without relying
on the denoise kwarg that newer comfy.samplers builds dropped
- baked settings: sampling_mode dual, upscale_by 1.0 (final size IS the
editor size), upscale_model as above; the engine's upscale-skip
condition now compares actual first-pass vs target dimensions instead
of upscale_by <= 1, which is behavior-identical for every existing
configuration
- segments are saved at the project resolution, so the Video Combine
restore passes through untouched - no RTX VSR anywhere
- MoxieProject refactor: component check and baked settings moved into
reusable helpers (_require_moxie_components, _baked_project_settings
classmethod) so the subclass composes cleanly; messages unchanged
- tests: Slow schema mirror, execute passthrough (sigma schedule math,
baked settings, context_frames forwarding, first-pass schedule), the
resolution-scale hook, and the denoise truncation helper
- docs: CHANGELOG entry, multitrack-schema.md note
MoxieProject subclasses EasyMultiTrackProject and runs the same
continuation engine, but its consolidated schema hid the new
context_frames widget added to the parent node in 724da63 - so the
overlap stayed locked at 22 for the node end users actually see.
- add the same context_frames combo (5/22/39/56, default 22) between
steps and seed on Moxie MultiTrack Project; execute() already
forwards kwargs through to the base engine untouched
- baked euler/beta sampling stack unchanged: sampling_plan stays
'custom', sampling_mode stays 'single'
- tests: MoxieProject schema order/options, execute passthrough proof
(widget reaches the base engine, baked settings never clobber it),
and the moxie_nodes schema assert updated for the new widget; test
harness gains a comfy.samplers stub and ControlAfterGenerate.randomize
- docs: multitrack-schema.md notes the widget on MoxieProject;
CHANGELOG updated
Default 22 keeps every existing MoxieProject workflow identical.
The number of frames copied from the previous segment as continuation
context (Context / Context Swap continuity) was hardcoded at 22. The
H3 context sampler nodes already accept 5 / 22 / 39 / 56, so the
project node now exposes the same choices as a context_frames combo:
- resolve the widget once in execute() and thread it through every
continuation site: the ContextSwap / MotionContextHard / HiResContinuity
graph nodes, the audio-lock prepend, the runtime latent trims, the
passthrough context encode, and the hires/low context re-encodes
- the context continuation now generates context_frames + 12 grid-tail
frames (was a hardcoded 34) so every supported value stays on the
H3 temporal grid; off-grid values raise a clear error up front
- tests: widget propagation at 5 frames end-to-end in the built graph,
off-grid rejection, schema order/options; existing 22-frame asserts
unchanged
- docs: multitrack-schema.md documents the widget; CHANGELOG entry
Default 22 keeps every existing workflow byte-identical.
The 3-step distill (speedup) LoRA is now baked into the diffusion model
itself, so the Moxie Multimedia Loader no longer applies any LoRA:
- moxie_loader.py: drop the models/loras scan, _auto_lora(), the module
level LoRA cache and _apply_lora(); the load chain is now
diffusion model -> attention backend -> CLIP -> VAEs. A leftover file
in models/loras is silently ignored and the folder can be deleted.
- moxie_project.py / moxie_preview.py: tooltips reference the baked-in
distill instead of the LoRA.
- tests + standalone loader script: assert no load_lora_for_models call
happens (the comfy.sd stub now raises if the loader tries), and fix
the pre-existing MMnode.tiny_vae stub gap (missing pack_vae_approx_files)
that broke the whole test module at fixture setup.
- README: models table drops the Turbo LoRA row; CHANGELOG entry added.
Node schema, outputs and sampling behavior are unchanged; workflows
saved before this change keep working.
- frontend/src + messages restored byte-identical to upstream v1.3.3:
format normalization shims, isMoxieEditorFormat gates, bundle marker,
tooltip reword, and their tests all removed; editor bundle rebuilt from
that pristine source so the editor behaves exactly like upstream
- format combo shows 'MiniMax' again (options=[MiniMax]); legacy queued
'Moxie Multimedia' values still normalized to MiniMax at execution
- T2V resolves natively again: t2v stays t2v in h3_task_type,
_multitrack_task_type and getMultiTrackTaskType; generation-mode t2v
mapping and the zero-media cloud fallback removed (upstream behavior);
Python tests and stub scripts updated accordingly
- editor delivery fix: core lets WEB_DIRECTORY overwrite the pack's
EXTENSION_WEB_DIRS entry, so the /extensions listing never contained
dist/release and a fresh browser never imported the editor (it only
ever appeared from stale browser cache). web/js/editor-loader.js now
imports the editor entry from the pack's own static mount, which
serves index.js with no-cache (chunks stay cacheable); the
/extensions/<pack>/index.js route was removed as a 404 trap
- shipped Moxie*.json templates carry format value MiniMax
Verified: frontend diff vs upstream = 0; bundle string checks + chunk
refs; vitest 565/566 (same pre-existing jsdom URL.revokeObjectURL env
failure); py_compile; test_t2v_r2v_and_loader.py ALL PASS;
test_rtx_upscale_integration.py ALL PASS
- T2V selections (editor default mode, explicit [start-end,t2v] override
tags, raw t2v values) now resolve to r2v in every task-type resolver
(h3_task_type, _multitrack_task_type, frontend getMultiTrackTaskType);
h3_generation_mode maps raw t2v to reference conditioning as belt and
braces. Reference-mode with zero reference media renders text-only
locally; the cloud path falls back to its native t2va request instead
of raising. i2v/l2v/v2v/edit flows unchanged.
- MoxieMultimediaLoader now presents a single model picker: LoRA, video
VAE, audio VAE, and CLIP widgets removed and auto-selected from the
pack's models/ folders (first sorted file, loud error when missing,
LoRA applied at strength 1.0 when bundled). Outputs and load order
unchanged; saved workflows keep working (extra widget values ignored).
Verified: py_compile; scripts/test_t2v_r2v_and_loader.py 50/50;
frontend vitest 566/567 (one pre-existing env failure on HEAD);
bun run build; scripts/test_rtx_upscale_integration.py all pass.
MoxieCombine now exposes exactly images / audio / fps - the leftover
VIDEO and FILENAME_PREFIX outputs are gone, so the output file name is
chosen on easy saveVideo itself. The base easy multitrackProjectVideoCombine
keeps its VIDEO and FILENAME_PREFIX outputs for legacy workflows.
The shipped Moxie-consolidated template is rewired accordingly: fps/images
links follow the new slot indices (2/0), the filename_prefix link is replaced
by the same project-scoped literal the removed output produced, and the
TEMPLATE's save location is unchanged.
pytest: new schema/execute/blocker coverage for MoxieCombine via a
_load_moxie_combine helper; base-combine assertions untouched.
MultiTrack and H3 conditioning cache logging printed Chinese labels
(ProjectMedia/SegmentMedia statuses, timing and cache-size breakdowns).
Translate every console-facing string to English so the ComfyUI log is
fully English-readable:
- cache status values are now 'cache-hit' / 'first-load' (was Chinese),
including the _all_media_caches_hit sentinel comparison and the
_easy_media_cache_status data written into tracks_info
- cache size breakdown now reads 'file(...)=conditioning(...)+...'
with English category labels
- timing/stage messages, video-track compositing and locked-audio
extraction messages, and llama.cpp install errors are English
- fullwidth '|' in the media_loading timing line replaced with ASCII
- test assertions updated to the new English strings
No behavioral changes beyond the log/status wording; the Chinese
prompt-reference syntax (@图片N etc.) and LLM prompt templates are
intentionally untouched as they are functional.
Newer ComfyUI builds lock registered v3 node classes after registration
(comfy_api internal __setattr__ guard), so _apply_lora writing
cls._loaded_lora raised: Cannot modify class attribute '_loaded_lora'
on locked class 'MoxieMultimediaLoaderClone'.
The (path, sd, metadata) cache now lives at module level, which the
locking mechanism does not touch. Test monkeypatch updated to match.
The Moxie model set is MiniMax H3-only, so no other model choices are
exposed in the editor:
- Moxie Multimedia Loader: removed the clip_type dropdown (which listed
every CLIPType recipe in ComfyUI). The H3 Qwen3-VL text encoder is now
always loaded with CLIPType.MINIMAX; older ComfyUI builds raise a
clear error instead of silently loading the wrong encoder.
- Moxie Multimedia Loader: outputs are now exactly model, clip,
video_vae, audio_vae - the legacy FAST_MODEL_LOADER (model_loader)
bundle output is removed. The model output keeps the attention
backend and LoRA applied.
- Moxie MultiTrack Project: takes model, clip, video_vae and audio_vae
as required direct inputs instead of a bundled model_loader; the
fallback plumbing is removed while EasyMultiTrackProject keeps its
model_loader input for non-Moxie workflows.
- Moxie Preview Override: dropped the model_loader pass-through slots
and patches the model input directly.
- Moxie-consolidated.json reference workflow updated to the new slots.
- Added tests/test_moxie_nodes.py covering the schema contracts, the
attention -> LoRA -> output chain, the MINIMAX lock and the project
loader synthesis.
The upstream locales/zh/nodeDefs.json is not part of this repository,
so every test reading it failed at runtime. Guard all 16 call sites
with pytest.skip and add a tests/conftest.py that stubs folder_paths
so collection works outside a full ComfyUI installation.