Krea 2 accepts reference latents, but they arrive with no position, so the
model treats them as a loose style hint: it borrows the look and reinvents
the content. That is the whole reason an outpaint drifts instead of
continuing the photo.
Krea 2 Encode attaches the reference latents and emits the negative from
the same node, so a turbo graph at CFG 1.0 stops carrying a second text
encode that does nothing. Krea 2 Outpaint Model Patch registers those
reference tokens into the target grid at the rectangle the stitcher
records, on frame axis 1.
Load Image + Pad gained a `reference` output - the unpadded source, fitted
to a 384px long edge and a multiple of 16, because the VAE downsamples by
8 and the DiT patchifies by 2. It is APPENDED after `stitcher`, never
inserted: a workflow stores links by output slot index, so any other
position silently rewires every saved graph.
The stitcher now records where the source sits, as source_bbox in pixels
and bbox_normalized in 0..1. Padding knows the rectangle exactly, so it
rides with the canvas instead of a parallel wire that can be left
unplugged. apply_stitch reads neither key, so an older stitcher stitches
identically.
Ported from the lab rather than copied: the lab's _krea2_helpers had its
own pad_image, feather_pad_mask, plan_pad_canvas, parse_fill_color and
build_canvas_stitcher. Duplicating the padding engine here would let the
outpaint path and the inpaint path disagree about what a stitcher is, so
the public module keeps only what is specific to Krea 2 and uses the
existing helpers for the rest.
Two departures from the lab version. The model patch declared a kv_cache
input and never read it - the cache was always on - so it is wired up now.
And it imports comfy's flux internals when you run it rather than at
startup: a core release that moves one then surfaces as an error on that
node, instead of deleting it from the menu and leaving saved workflows
reporting it missing.
The old display name led with the mechanism. "Image Compare A/B" says
what the node is first, which is what someone scanning the menu reads.
"image compare" joins the search aliases; "compare" and "a/b" stay.
The mapping key AUSBOSS_NODES_Compare is untouched - it is the
workflow-compatibility contract - so saved workflows are unaffected.
Load Image + Pad does the same job from the same stage and starts from
the file, so the only thing the second node added was taking an IMAGE
wire - one idea holding two slots in the menu.
Deletes the node, its panel, its help doc and its stage-only shared
helpers (padEmptyStateText, findPadPreview), and drops it from
NODE_MODULES, PUBLIC_NODE_IDS, the README and the panel guard list.
The padding helpers are untouched; Load Image + Pad and Align Image both
use them. Two console labels there named the removed node, so
_fill_tensor now takes the source name - Align Image passes its own
pad_color, and a colour it cannot read is no longer reported against
padding. The BHWC and mode errors read "Padding"/"Pad mode" instead.
The on-node pad stage moves to ComfyUI-AusBoss-Lab, which has its own
AUSBOSS_LAB_PadImage.
The node produced a mask offset, grown and blurred - no contact
darkening, no perspective - so it read as a sticker halo rather than a
shadow in the scene. That is the approach, not a tuning problem, so it
comes out rather than shipping under the pack's name.
Deletes the node, its helper module, its test file and its help doc, and
drops it from NODE_MODULES, PUBLIC_NODE_IDS and the README. The colour
helper tests keep their two-callers case using Pad Image and Transform,
the real remaining call sites.
blur_mask and grow_shrink_mask stay in _mask_helpers - Crop For Inpaint
and Pad Image use both.
The audit compared every node against ComfyUI core and the major packs;
this lands its outcomes in one branch.
On-node visual editing (the old Codex-pack interaction, rebuilt):
- New Load Image + Pad: image loader with whole-edge pad dragging on the
node canvas, align-to-multiple rounding, and a megapixel target that
rescales the source before padding so the mask seam stays binary.
- Pad Image gains the same handle canvas, fed by an execution-time
snapshot of its input; wireframe with diagnostic empty states before
the first run. Widgets, outputs, and saved workflows unchanged.
- Image Crop + Rotate + Pad's compact panel is now a live stage - crop,
pad, and rotate handles on the node, one shared implementation with
the full-screen editor so the two surfaces cannot drift.
Feature gaps closed against the ecosystem:
- Crop For Inpaint: target_megapixels, selectable rescale_algorithm
(recorded in the stitcher for the paste-back), and extend_* inputs
that grow the frame itself for outpainting.
- Save Video: format choice - mp4 h264, mp4 h265, webm vp9 with Opus
(resampled to 48 kHz); CRF drives all three.
- Load Video: every_nth thinning (fps output divides to match) and a
max_frames cap that ends the decode early.
- LoRA Loader: saved stack templates behind a bar button, and a console
warning when a LoRA's base-model family disagrees with the checkpoint.
- LM Studio Chat: history in/out for multi-turn chains and a json_schema
widget for structured replies.
- Drop Shadow: multiply blend and a shadow_mask output. Align Image:
offset outputs for un-aligning after sampling. Color Match:
first_frame reference mode as a one-node video flicker fix.
Removed (backed up in git history): Select Frame Range and the Video
Bundle family - core ImageFromBatch and the core VIDEO wire now cover
both jobs. Dead NODE_ID module variables cleaned out of every node file.
455 Python tests, 216 JS tests, validate_nodes.py green on 19 node files.
Every display name swaps ' (AusBoss)' for ' 🆎' - menus, docs, tooltips,
the README, and the example workflow's notes. Mapping keys never change,
so saved workflows keep loading; 'ausboss' keeps finding everything via
the category, the AUSBOSS_NODES_ prefix, and SEARCH_ALIASES. Python error
messages drop the suffix entirely so exception text stays ASCII for
cp1252 consoles.
The LM Studio endpoint toolbar declared no getMinHeight, so the frontend
gave its wrapper a 12px sliver and the buttons clipped; it now declares
32px through every sizing path, and the buttons breathe at 22px.
Align Image snaps width and height to a clean multiple - 16, 32,
whatever the model wants - because Qwen image models, VAEs and plenty of
samplers only behave on cleanly divisible sizes. Each mode implies its
own rounding, so there is no separate direction widget: resize rescales
to the nearest multiple, crop center-crops down, pad replicate-pads up,
and a side smaller than one multiple always grows to exactly one rather
than erroring. The new size comes out as INTs for wiring into latents.
Image Size reads width, height, longest edge and shortest edge off the
wire so resizes and latents key off the actual image instead of retyped
numbers that drift.
LM Studio Chat posts a prompt - and optionally one image - to a local
LM Studio server, or anything else speaking the OpenAI-compatible chat
API. The interface is the middle ground the installed ecosystem packs
miss on either side: endpoint string, model string where EMPTY means
whatever the server has loaded (a dropdown fetched at ComfyUI startup
goes stale whenever LM Studio was not running yet), optional system
prompt, prompt required unless an image is connected, temperature,
max_tokens, seed as the re-roll knob, a timeout, and an image downscale
cap. Reasoning-model think blocks land on their own output so the text
stays clean for conditioning. Stdlib urllib only - no new dependencies -
and the POST runs off the executor thread. Every error names what to
fix, with the server's own words when it had any.
Verified live against a running LM Studio: a JIT-loaded qwen3.5-9b
answered a text round trip in 7.6 s through the real node; the
no-model-loaded 400 and a dead port both surfaced their actionable
messages. The vision payload shape is unit-tested; the two local VL
quants refused images server-side (missing projector / failed load), so
the live vision check stays on the deployment checklist.
48 offline tests cover the alignment math, both size nodes, the URL and
payload builders, think-splitting, image encoding, and every mocked
network error path.
Every registered node now has rich help: Video Unbundle and Video Bundle
Edit only existed inside the family page, which ComfyUI's help panel
never shows because it looks docs up by exact node id, and the
SimpleWatermarkRemover compatibility alias had nothing pointing users at
LaMa Inpaint. 21 of 21 mapping keys now have a js/docs page.
Three statements the review verifiers caught as false stay fixed too:
the README said no minimum ComfyUI version is declared while
pyproject.toml declares and Manager enforces 0.27.1; the optional-extras
comment promised a warn-and-fall-back that only Stitch Inpaint actually
does, where Refine Mask deliberately stops with the package name; and
the Load Video tooltip cited 0.3.26 for a VIDEO type that landed in
0.3.31 - moot either way, since every supported core has it.
Frame Chooser, mostly. A pause did not survive a workflow load: LiteGraph
clears a graph by removing every node, so undo, a tab switch, Clear
Workflow and opening another file all fired the teardown a deleted node
uses to release its pause, and killed a run that was still going. That
now only fires for a real deletion. Answering "keep all" wrote the batch
out as 1,2,...,N into pick_list, so the next run dropped whatever the
enumeration did not reach or failed outright on a shorter clip; it
writes the empty answer it was given. And pick_list is only written back
under "keep last selection" - it pre-answers the node, so filling it in
automatically meant a chooser left at the default paused exactly once.
Refine Mask: smooth applies only the change it made to the jaggy edge
rather than binarizing everything, so an already-soft matte keeps its
alpha (binary masks stay bit-identical), and blur now reaches the
matting solve instead of being thresholded back out of the trimap.
Frame Interpolate's copy path honours batch_size; it gathered every
copied frame at once, which on a long clip at an integer multiple was a
multi-gigabyte allocation on the frames device that no setting bounded.
Subgraph execution ids are resolved by walking the chain instead of
stripping the prefix, which could match an unrelated node with the same
number. Status and runtime badges appear inside subgraphs again, and a
subgraph pause renders on the node that actually paused.
Also: the Custom colour scheme no longer strips every node's colour when
its stored colour is unreadable; Drop Shadow and Pad Image name their
own widget for an unreadable colour instead of blaming Transform; a
non-ASCII token and a non-object JSON body answer 400 rather than
throwing a 500 and stranding the pause; pick_list rejects Unicode
digits, which used to raise during validation or silently read as a
frame nobody typed.
Tooling: validate_nodes parses NODE_MODULES rather than matching it
anywhere in the text, and an unregistered module is an error - it used
to exit 0 printing nothing. It also refuses mapping keys declared
outside nodes/, which is how the registry fixtures came to be advertised
to ComfyUI-Manager as installable nodes; they are .py.txt now. And the
video test no longer leaves ComfyUI on sys.path, which was flipping
every fail-soft seam for whatever ran next.
Continue, cancel and the expiring countdown were all writing straight to
the pending object and reporting success. Two clicks a frame apart both
returned 200 with the second selection winning; a cancel behind a keep
turned a continuing run into an interrupt; the timer could discard an
answer the browser had just been told had landed.
Every terminal path now goes through claim_pause, which records the one
outcome under the store lock and tells every later claim it lost. The
route answers 410 for those, the panel treats 409/410 as already-handled
rather than an answer failure, and a resolved pause stops being offered
to a page that reloads before its waiter unwinds. A rejected selection
claims nothing, so the cancel a deleted node posts behind it still
releases the graph.
tests/test_chooser_pause.py drives the races with real threads; ten of
its nineteen tests fail against the previous unguarded write.
The pymatting solve costs roughly 90 ms per megapixel of paste window,
per frame - 48 ms for the 768x768 window a 1024x1024 frame produces,
106 ms for the 1440x816 window from 1080p, measured on a 16-thread
desktop CPU. A video batch with the toggle on therefore ran to the end
with no feedback and no way out. The spread now checks for a cancel and
reports progress between frames, through the same seams the LaMa helper
uses, and stops at the next frame boundary.
It also hands pymatting float32 instead of float64. The estimator casts
whatever it is given down to float32 before solving, so the wider
temporary bought nothing: outputs stay bitwise identical across the
toggle, batch and rescale paths, and writing frames into one buffer
instead of stacking them cuts peak memory growth by about a third on a
48-frame 1024x1024 batch.
Node help now carries the measured per-frame cost and says plainly that
the toggle is for finishing a chosen take, not for a long exploratory
batch.
Keep selected, Keep all, Cancel, Enter, Escape and the cancel a deleted
node sends all posted without checking whether an answer was already in
flight, so two quick Enters put the same one-use token on the wire twice
and the duplicate's 409 landed as "Answer failed" over the success the
first request had already painted. Every path now goes through a latch on
the panel state: a second submit never fires, the controls read as
unavailable until the reply lands, and the latch clears on success or
failure so a genuine network error can still be retried. Each submission
also captures the pause it was posted for, and a reply is dropped when the
panel has since resolved or moved to another token - which covers both the
late rejection and a new pause arriving on the same node mid-flight. The
latch and the staleness predicate live in js/shared/frame_chooser.mjs with
node:test coverage.
A feathered paste re-blends the background that the sampler already
mixed into its soft edge, which rims the seam. The new fix_edge_halo
toggle estimates the unmixed foreground color with pymatting and pastes
that instead, spreading on a slightly grown blend mask while
compositing with the ungrown one. The toggle is appended and off by
default; without pymatting the node warns once and pastes unchanged.
Pixels outside the blend region stay bit-identical either way.
A pause that starts out of sight - hidden browser tab, node scrolled off
the canvas, or another workflow or subgraph fronted - now raises one
toast naming the node and pulses the waiting workflow tab when the
frontend exposes it. A filmstrip the user can see stays quiet, and the
notice latches on the pause token so re-renders never repeat it; the
check runs a beat after the pause lands, once the canvas has placed the
widget. The focused panel takes 1-9 to toggle a frame, A and N to flip
the whole selection, Enter to keep, and Escape to cancel, consuming only
mapped keys and leaving anything typed into a field alone. Clicks inside
the panel are ignored for 400ms after it appears, so a click aimed at
the canvas cannot answer a pause that just popped up. The decidable
parts - key map, cooldown, visibility - live in js/shared with
node:test coverage.
Covers smooth, black/white point, edge_refine, and guide_image; spells
out the fixed op order and which controls need the optional
opencv-contrib-python / pymatting installs versus the stock install.
A filled pick_list ('1,4,9', commas or spaces) applies immediately with
route-grade validation: no pause, no thumbnails, no events, and a typo
fails the run loudly. Answering a pause now writes the kept indices back
into the visible pick_list widget over the done event, so the next queue
reproduces the choice headlessly in any tab. IS_CHANGED returns a stable
canonical fingerprint while pick_list is filled, letting pre-answered
runs cache; empty keeps the per-queue re-pause token.
A reload drops the DOM panel while the server keeps waiting. The pause
announcement payload now lives on the pending record, and a new GET
/ausboss/frame_chooser/pending route re-serves it with the live
countdown remaining. The frontend queries the route once the graph
configures (plus a startup fallback) and re-renders the filmstrip for
any chooser still waiting, leaving already-active panels untouched.
New optional widgets timeout_seconds (0 = wait forever) and on_timeout
(keep all / keep first / keep last / cancel). The blocking wait arms a
monotonic deadline, emits a once-per-second tick event the panel renders
as a countdown in the header, and applies the policy on expiry; cancel
raises the same interrupt as pressing stop. A done event releases the
panel in every open tab when the timer answers.
Pads any side with a solid color, an averaged edge-color band,
replicated edge pixels, or a blurred and dimmed pillarbox backdrop.
The original region stays bit-identical in every mode and the second
output masks exactly the new padding, ready for outpainting.
Per-channel LAB mean/std matching against a reference image, with an
optional mask that restricts both measurement and application, a
strength blend, and batch-1 reference broadcasting. Pixels outside the
mask stay bit-identical.
An OUTPUT_NODE that saves temp previews of each batch's first frame
and renders them stacked in a DOM panel: slide mode clips B to the
pointer with a seam line, hold mode flashes B while pressed. image_a
passes through unchanged so the comparison can sit mid-graph. Pure
clip and mode logic lives in js/shared/compare.mjs under node:test.
Video Bundle packs frames, fps, and optional audio into one custom
wire; Video Unbundle unpacks it; Video Bundle Edit copies a bundle
overriding only the connected inputs. Frame count, size, and duration
are derived from the frames tensor and fps at bundle time and
recomputed on every edit, so they can never drift from the frames.
ausboss