The H3 run trains with a loop of our own (LoRA on the attention and feed-forward projections of the
50 blocks, per-block checkpoint, uniform noise levels, musubi-tuner's approach) while ComfyUI carries
the batches. Every Log record is flushed to disk as it is said and reloaded on start; the run
writes each stage and step to it. ComfyUI's own offloading toggle is ignored by the new loop.
UNTESTED on the real weights.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
Measured on a tiny H3: backward growth was ~75 MB per block, now flat (61-88 MB at 2, 8 and 16 blocks).
The in-place mul_/addcmul_ on slices of the stream in the block helpers were the cause.
UNTESTED on the real weights.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
Block swap keeps the frozen weights of the last N transformer blocks in system
memory and lets ComfyUI's weight casting fetch them one layer at a time. The idea
and the block count as the one knob follow musubi-tuner's block swap (kohya-ss,
Apache License 2.0); this is an independent implementation.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpnF7GwNGMbngU11wjym7P
Prompt Markup strips it before the encoder; Apply Prompt Markup puts weights
and timed windows on attention and blends on the conditioning. Where the words
sit is the H3 module's (text_tokenizer, prompt_rows, video_time_rows). Wired
into both H3 presets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
H3's latent upscaler loads through the H3 module (ComfyUI's loader has no
branch for it); an upscale resizes H3's anchor pin so pass 2 stays pinned.
FunPack Sampler takes an optional wired schedule, run exactly as written.
New preset: H3 Reference to Video + second pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The H3 module offers anchor_pin, so ALG blurs the pin's latent without
learning H3's payload layout.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Block repeat's video-only mode now really keeps audio single-pass: H3 blocks
write in place, so v4 restored audio from an already-repeated stream.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The port plan calls for every module under modules/ to declare a
CATEGORY so Settings/Engine panels can group by it. Only alg, sharpen,
appearance, and full_control had one; the other 12 (loaders, empty
latent, media load, minimax_h3, output decode, sampling modifiers/
sampler, system pipeline) are plain infrastructure with no user-facing
settings dial, so they land in "system" except output_decode (post,
matching its STAGE) and the two sampling modules (sampling). Verified
against core/tests (548 passed) -- category validation and the
registry's field-mapping catch a typo here, per the earlier fixture-gap
fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
keep_proportion/extra_padding/invert_crop (ImageTransformKJ) and format
(SaveVideo) were wrapped in a redundant {"field": "value"} dict instead of
the plain string ComfyUI's own DynamicCombo expansion expects at the wire
level -- every one of them silently failed graph validation and fell back
to the node's own default, undermining the "total_pixels" resolution-fit
behavior this preset exists for. Found by loading the preset in a real
browser against real node schemas; the dev server never had comfy_extras
or KJNodes registered before now, so this was invisible locally until
tools/devserver.py also gained real node registration (init_extra_nodes +
KJNodes' load_custom_node, matching what the pytest `registered` fixture
already did).
The two pipeline_preset tests that were supposed to catch this had too
broad a filter -- it excluded ANY "is not one of" message as an assumed
"combo not chosen yet" state, which in this codebase actually shows up as
"needs X and nothing fills it" instead. Tightened to exempt only the three
loader-file combos genuinely left unset on a fresh install.
Verified live: preset loads with zero combo-validation problems, only the
three expected unset-loader-file messages remain, and Generate correctly
refuses with an honest reason rather than a false success.
Pieces B and C: Scene gains source_image (aspect-ratio reference only --
resolution itself stays a project.video setting, ImageTransformKJ's
"total_pixels" keep_proportion fits the picture's aspect into that budget)
and references (media ids, in order). Inspector's Scene tab gets pick/clear/
remove controls; picking routes through the existing Media bin via a small
pending-pick handshake in boot.js (one bin, one onPick, so "which field
asked" has to be tracked between the click that asked and the click that
answered).
References wire into MiniMaxH3ReferenceToVideo's ref_images autogrow input
dynamically, per run, only for what a scene actually has -- an unwired slot,
not an empty image, is what "not in use" means to that node. The target is
data on the role itself (wireTo), not a slot id boot.js would otherwise have
to know, keeping the "core never names an implementation" contract intact
one layer up into the app.
Verifying this exposed a second real core gap (after the MatchType one):
core/graph.py had no idea what an Autogrow (V3's io.Autogrow) input even
is -- "ref_images" reported as one opaque COMFY_AUTOGROW_V3 slot, its real
per-instance sockets (ref_image_0, ref_image_1, ...) invisible to
declared()/build(), so a wire naming one was refused as an input the node
does not have. Fixed generically in comfy_types.py (autogrow_instances())
and graph.py, confirmed against ComfyUI's own _io.py (finalize_prefix's
dotted join, build_nested_inputs's runtime un-nesting) -- not guessed, and
proven end to end with a real override()+build() test using the exact edit
boot.js makes.
