A Flash exam showed a light model can build a route it is told, so the reminder
no longer sends the user to a flagship model: the route is untried, see how the
run turns out, and save a card if it is good.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three exams with Gemini 3.8 Flash on an isolated ComfyUI all went the right
way. They showed that a base kept by Anomalous was reported as a file changed
outside it, and that the one-pass route still told the AI to show bases
first. begin_task's words are shorter: with a real node list it went past
its 4000 characters.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A node's "Mark this state" keeps the serialized canvas as a mark in a bar
under it. A mark's number goes back to that state, each has its runs, and
seeds are Random or Locked (one per round, shared by the marks; a fixed
seed alone would come from ComfyUI's cache). Run checks every mark with
validate_prompt first, then compiles detached copies with their seeds
written, so the canvas and its seeds stay as they are.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With two main models picked, ask_user(routes=true) shows the routes this
ComfyUI can build for them, each with what it does to the picture and a
sample, a saved card using both first; the answer says what to call, and
add_nodes(route=...) builds the route from the ways on a new tab, the LoRAs
on each model. get_toolbox lists each model's cards as its examples. A way
counts as ready when a family variant has what it needs. An AI placing
nodes itself is a new route: one reminder a task, or with the new gate the
user's OK.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
add_nodes(use="picture") builds a text-to-image graph on the model the
user picked for the task, or add_nodes(model=...): one checkpoint for SDXL
and its kin; for Anima the UNET with its text encoder and VAE, found in
the model folders ("files"), and the tested base-pass sampling. A
standalone block is built only on a canvas making no picture yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upscale held by Tile takes in the tile upscale, the Anima Tile detail
upscale and the second-model upscale: its redraw has its own copy of the
canvas's model line (copy_of "$sampler.model": loader and LoRAs, titled
"Upscale model"), so a second model is one set_model and the LoRAs stay;
on Anima it is the user's tested route (latent 1.22x, Tile LLLite,
denoise 0.85). The iterative upscale goes. add_nodes no longer lists every
block name, and the repaint hints name the use, not a block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
list_workflows gives each card a route read from its saved graph: its
passes in run order, the model each draws with, a split's steps, an
enlargement, control models such as a Tile LLLite and what comes after.
A Flash run picked a one-pass card for a two-model job; it saw names only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_workflow(base_only) queues the first pass alone (base_only.js cuts each
queued prompt; save nodes become previews, the canvas stays). When it ends,
ai_bases.py shows the previews on a picks panel; the picked one is copied
into the output folder and its seed held. The next run takes the kept
picture where the rest takes the base's picture, so the base is not drawn
again; a base handed on as a latent (mid-way switch) is drawn again from
its seed and compared with the kept one. Drawn again from the same seed, a
base once came out grainy on a 6 GB card, hence the kept picture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hand fix's "3 tries in 4" holds for the pictures tried, not for crowds or hard poses;
its checks now compare with the enlarged redraw and look for seams, missed hands and good
hands made worse. Enlarging at the mid-way switch asks the user first when the workflow
already enlarges the picture downstream. Both ways record GPT's runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hand-fix way now redraws each hand SAM3 finds by the word ("hand:8": the word alone
takes only the likeliest hand) inside its own outline before the hand detector's pass, as
the reference author does. The pass is optional: its nodes go when SAM3's node or model
(sam3.1, now in the ways' model list) is missing, with the nodes that only fed them.
Switch models mid-way says how to enlarge at the switch: a LatentUpscaleBy after the first
part, which then finishes its picture, and the second part adding noise anew. Enlarging
the leftover noise itself (the author's wiring) covered the picture in coloured blocks on
Anima at step 21 and 12. The staged-picture skill says so in one sentence.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From a picture check of the staged ways (another AI judged the pictures):
- The staged skill shows the base first so the user can pick it. With
the sampler's seed moving after each run, adding the second pass and
running again redrew a different base. The skill now says to hold the
base's seed (control_after_generate fixed) before that first run.
- Split er_sde came out grainy on two Anima models but cleanly on Nova
Anime XL (one seed each). The mid-way switch no longer says er_sde
never splits: it keeps dpmpp_2m_sde_gpu on Anima and calls it the
safe choice elsewhere. Its SDXL version ran for the first time (Nova
Anime XL to MiaoMiao Harem v2.0, step 18 of 30, 41 s on 6 GB) and is
recorded as tested.
No values changed. Full suite and structure check pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3d720ef made mcp_fill.saved_workflow return the set too, and
run_workflow took it, but open_workflow still unpacked two values and
failed on every saved workflow ("too many values to unpack"). It now
passes the set on to the page like run_workflow, so "workflow / set"
opens with that set applied.
A test calls open_workflow with a workflow and with one of its sets; it
fails with the old line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The skill staged_picture says how, not values: different models for a
picture's early and late steps (refine midway); a locked redraw while
enlarging, which keeps the model or switches in that same redraw to a
second model of the same family (upscale detail, anima tile, steady),
with no extra refine pass after it; model after model (refine). It says
which switches need one family (any reusing the first stage's prompts and
VAE; the single-file refine with its own loader may cross families, its
prompts rewritten), and that a style LoRA the user picked goes onto the
second model when the look is to be kept. The user's saved cards and
combos come first; nothing the user did not pick is added.
Written to the user's handoff (.tmp, three points). Code checks only, as
asked: no picture was made.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New way switch_midway (refine, want midway): the canvas's sampler becomes
two KSampler Advanced parts, the first on the canvas's model up to step
21, the second on a new loader from there, taking the old sampler's place
and links. On Anima it takes the tested sampling (35 steps,
dpmpp_2m_sde_gpu, beta): er_sde comes out grainy when split, even with
the same model in both parts, which the AI is told.
refine_other_model and upscale_two_models get Anima variants: a UNET
loader "like" the canvas's, reusing its text encoder, VAE and prompts; on
Anima the second-model upscale is the Tile route the user verified (1.22x
latent, LLLite 0.7, denoise 0.85, ColorMatch).
A node "like" or replacing a sampler now takes its seed as noise_seed on
KSampler Advanced and the other way round. The scene argument's
description is shorter to keep the tool list within its budget.
Ran on the isolated test ComfyUI on an Anima canvas: each way built
through add_nodes and made a picture (79 s, 148 s, 127 s on 6 GB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A workflow card's saved parameter sets (方案) were the user's alone. Now
list_workflows lists each card's sets as "workflow / set" (those unlike
the card's own values, as the Workflows page lists them), and
run_workflow(workflow="workflow / set") opens the workflow and applies
the set with the page's own Apply code before filling in the blanks; the
tab is marked with the set too, so another set opens afresh. No new
argument: the tool list stays within its budget.
The Anima skill points at the sets: the daily run is the user's saved
workflow or one of its sets.
Checked on an isolated ComfyUI: a card built from the verified two-pass
Anima graph with its own values as the plain route and two sets (upscale
on another model, model switch mid-way); run_workflow on the second set
queued exactly its values and made pictures pixel-identical to that
route's verified ones. Full suite and structure check pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A skill now says how to judge and names the parts; models and values
live in the way and in the user's own combos and workflows. The Anima
skill drops the tested models, LoRA and sampler numbers (the
upscale_anima_tile way keeps its second-pass values), says to use a
workflow or combo the user saved for this route as it is, and keeps the
findings as judgment: show the base first, what Tile does and does not
keep, and when changing models between passes is untried.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AMB only uses: training goes to its own plugin. This takes back the
record of what works (api/ai_evidence.py), the training network and
rounds (api/ai_training.py, ui_ai_network*.js, the round form), the
Gates' record switch, the ways' order by record and the external_skill
field of log_outside that only fed the record.
Kept, as part of using AMB:
- the user's word under a finished run, now only told to the AI
(ai_run_control.set_verdict, ui_ai_verdict.js);
- the Gates' "Custom nodes: only the ways'" rule;
- a picks block left to the AI takes its recommended picture.
Reverts 73518dc, 67bf5b2, 6647b05 and 49d05cf, then puts those three
back with tests in test_ai_run_watch, test_ai_gates and test_ai_panels.
Full suite and structure check pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The side switch's training network now opens over the whole AI
workbench: numbers on top, a filter by kind, a graph or a table with a
cell per pair, and a detail column listing each run of a pair (model,
AI app, judgement, pictures, task), a base model's models and the
training rounds. New GET /anomalous/ai_training/runs; the network data
adds each model's runs and the rounds.
Not yet verified in the browser: Esc after routing it through the
browser's back buttons. Full test run pending.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The AI goes by the ready-made ways and skills alone by default; the Gates
window's "training network record" shows it the record, and a training
round the user starts always does. add_nodes still picks the way that did
best either way. The network view says whether the AI sees it now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The AI workbench's side column gets a training network view: base models
on the left, ways, skills and node packs on the right, a line for each
pair that ran, coloured by how it turned out (the user's word first, else
the AI's, dashed) and as thick as it ran often. A node or line shows its
record, last pictures and the pairs not tried yet in a card kept in sight.
"Start a training round" picks a skill (the AI's prompts with its advice
and without) or the ways of one use, up to four models and the pictures
each, and makes a task for the next AI (api/ai_training.py). Each run that
ends fills the trial it fits by what it used; the AI is told the next trial
and, once a model's trials are done, to compare them in a picks panel with
its own pick marked recommended. The record keeps the AI's pick apart from
the user's, and a picks block left to the AI takes its pick.
Tests: tests/test_ai_training.py; full suite and structure check pass.
Checked on an isolated ComfyUI: a round started from the view, four real
runs matched to their trials with the next one told each time, the
comparison panel and the user's pick reached the record and the graph.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each run the AI starts now counts for the ways it still has, the skills
fetched before it and the custom node packs of its nodes, by the base
model of its main model (api/ai_evidence.py). How it went comes from what
the user does anyway: a picture picked in a picks panel or kept in the
gallery, one passed over, or their word under the run (Turned out well /
Not quite, told to the AI as their message).
begin_task and get_toolbox give the AI the record in a few words, add_nodes
a way's or the packs' own, and add_nodes(use=...) picks the use's way that
did best on the canvas's model. The Ways panel shows each way's and skill's
record by model. The Gates window gets "Custom nodes: Any / Only the ways'",
which keeps the AI to the ways' nodes and the task's picks.
Tests: tests/test_ai_evidence.py; full suite and structure check pass.
Checked on an isolated ComfyUI: a real run was recorded with its model and
skill, the verdict reached the AI, and the ways-only rule refused a custom node.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ai_skills.get refuses a private-only skill in a task that is not private and every other
skill in a private one, so a private task's AI gets the privacy skill alone; get_toolbox
tells it by the task. The Ways panel lists the skills of the task's kind and is drawn
again when the task's lock is turned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ai_sealed keeps a private task's prompt text and picture file names behind tokens ([P1],
[img1], task["sealed"]): mcp_server puts the texts back into a call's arguments for the tool
only, and ai_privacy.guard has every result masked on its way out (a node's STRING inputs
by ComfyUI's node list, prompt fields, picture names, any sealed text, also in the text
content). The user's own canvas changes are told sealed. A "prompts" panel block is the
user's prompt box (ui_ai_prompt_box.js): prompts, a negative one and reference pictures from
the gallery or dropped in, answered only from Anomalous, the AI getting tokens; above it the
prompt_pose skill's words for the canvas's model. A new private-only skill, privacy, tells
the AI how to work, and begin_task points a private task at it. Tests plant canary prompts
and picture names in every tool's result and check none gets out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each task has a lock in its header (task["privacy"]); the Gates window sets the default for
new tasks, and turning the default off pins the tasks it covered. mcp_server hands every
tool result to ai_privacy.guard last, after run pictures and the user's messages were
added, so a private task's image content never leaves; a call tied to no task is private
while the default is on or a private task is unfinished. The card in the AI app shows no
pictures for it, begin_task tells the AI, and turning it on or off is a user step and a
decision. A test calls every tool with a picture in its result and checks none gets out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Checked against the headers of 315 models in ComfyUI-Manager's list: a Hunyuan-DiT
checkpoint and LBM were read as VAEs, SUPIR as a ControlNet, XLabs' Flux LoRAs and
IP-Adapter as diffusion models, a whole CLIP as a text encoder only; also PhotoMaker, LoRAs
of weight differences or OFT, and checkpoints without a VAE. Files two loaders read from
different folders (SUPIR, a whole CLIP) are not sure. A lone drawing model in checkpoints
loads there (the checkpoint loader reads it as a diffusion model). Newly told: diffusers
UNets, Stable Cascade, SDXL T2I-Adapters, RCAN/SPAN/compact upscalers. 301 of 315 now agree.
"Looks like no drawing model" said what it could not tell: now "Kind not told". Kept header
answers carry the rules version and are read again when it changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An Anima LLLite in controlnet is not misplaced while the ComfyUI-Anima-LLLite pack
(which reads it from there) is loaded, and identical copies kept one per loader folder
(controlnet and model_patches) are listed as shared, both in use, not as duplicates to
recycle. An IP-Adapter carrying LoRA layers (the Anima IP-Adapter) is no longer read as a
LoRA. The header cache keeps the new "form" field (older entries are read again).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
prompts (1-8, instead of prompt; at most 16 tries with batch) fills and queues each prompt
in turn with the call's other blanks; the receipts form one run and by_prompt says which
are which. A prompt that cannot be filled or queued stops the rest, the queued ones stay
tracked and submission_error names where it stopped. The run_workflow description is
tightened so the tool list stays within its budget.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each skill the AI fetches with get_toolbox(skill=...) is a "skill" step of its task
(ai_skills.note_use), with the form its advice came in for the canvas's model; the same
advice again before the next run adds nothing. A run's row lists the skills fetched since
the run before it (ai_task_view.skillsBefore). A skill step never names an untitled task.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The prompt_pose skill now comes in the form for the model the canvas's sampler uses:
its drawing model's parameters are counted from the file's header (model_kind.
diffusion_params: safetensors shapes, GGUF tensor dims; a checkpoint's text encoders
and VAE left out), so a compressed GGUF counts like the full model. From 10 billion
(Flux.1, Flux.2, Qwen-Image) the AI is told hands and hard poses come out well and
not to hide or simplify; under it, the clear-pose and optional hand-hiding advice;
when the model cannot be told, both cases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Files a scan found on the same Civitai model page (metadata model_id; never a similar
file name) share one card in the model list, the newest version shown or the one picked
last (kept in this browser). The card's version label opens the shared action menu to
switch. The details list the versions here under the name, and "All versions on Civitai"
(GET /anomalous/download/versions, this computer only, asked only when pressed) lists
the page's versions to open or download beside this one, checked by SHA-256, a taken
file name getting the version's name added.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The pose and hands advice for writing prompts moves out of the set_prompt and
run_workflow descriptions into the first skill, prompt_pose (api/skills/). begin_task
lists the skills that are on in a line; get_toolbox(skill=...) gives one; the prompt
arguments only point at it. The tool list is back within its budget (it was 44726 of
44000 characters). The Ways panel lists the skills under the ways, each with a switch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.5x left the Codex test picture's hand better but still soft; 2x gave
sharper fingers in the same seeds (about 105 s instead of 75 s on a 6 GB
card). The fix-hands way now enlarges to about 4.2 megapixels, never
smaller, at most 2x: reference_blocks.enlargement reads the picture's
size (add_nodes passes picture_size of its input copy). Tried through
AMB on an isolated ComfyUI: 2547x1656, no memory trouble, 2 clean hands
of 4.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The comfortable size was 4.2 megapixels on 12 GB, 2.1 on 6 GB: the fix-
hands way's 4-megapixel redraw got "ask the user whether to stop" though
a 6 GB RTX 4050 ran it (Anima, 10 steps) in about 100 s. It is now 8.4 on
12 GB (4.2 on 6 GB); the decode warning keeps its old line in its own
constant.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A block (detail, want hands) that takes a picture: enlarged 1.5x and
redrawn lightly by the canvas's sampler (denoise 0.4), then every hand
found by hand_yolov8 redrawn at guide 768, denoise 0.45 ("detailed
hand"), saved as hands_fixed. Tried on the Codex test picture on an RTX
4050: the hand pass alone fixed about half the seeds, enlarging first
about 3 in 4; the way built through AMB on an isolated ComfyUI gave 2-3
good hands of 4.
On an Anima model a variant adds the Anima Turbo LoRA (0.7) for 10 steps
at CFG 1, about 75 s; a variant's LoRA missing falls back to the plain
block with a word on what it would do. Its next step tells the AI to
make 4 tries, check the hands itself, retry at most twice, and let the
user pick in a zoomed panel. The Turbo LoRA joins the recommended models.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a video's first frame could not be read (a broken file, no decoder)
the card's thumbnail route served the whole video as the picture. It now
gets the same SVG placeholder with its size as a picture with no card.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex's hand tests (2026-10-29): at denoise 0.55-0.75 the repainted hand
filled its whole mark and came out too big, at 0.4 it kept its size. The
crop-and-stitch block now repaints at 0.45 (the built-in one at 0.5), and
its next step and the inpaint note tell the AI to mark tightly, keep the
strength low, make 4 tries and let the user pick.
Slider and picks blocks take a zoom {x, y, w, h}: every picture shows
only that part, enlarged, from the picture itself. The checks of block
parts move to ai_panel_parts.py (ai_panels.py was at the 600-line line).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_workflow's image filled a LoadImage with "name.png [output]"; once
ComfyUI's run history is gone, its newer frontend marks that value as a
missing file though LoadImage reads it (Codex found this on 2026-10-29).
An output picture is now copied into the input folder (content-named,
transparency kept) as the gallery's menu already does.
get_situation's canvas errors also carry what the page's missingMedia
store marks missing, so the AI sees the red error the user sees.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_workflow's batch queued the same prompt N times when the sampler's
seed was held "fixed": ComfyUI ran it once and the user had one picture
to pick from. While a batch is queued a fixed seed counts up instead
(the native control moves it per prompt); afterwards the control and the
seed are as they were.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The run estimate followed each sampler's size but not what a SaveImage
finally gets: a 5120x3328 sample upscaled 4x saved 20480x13312 (324 MB)
without a word. A save past 64 megapixels now warns on the timeline and
in the AI's run answer, with the PNG size it will take.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A 20480x13312 upscale is past Pillow's pixel limit, so its card image
failed and the catch-all fallback served the 324 MB original as the
thumbnail: the gallery and the AI workbench decoded it every time.
- A picture past the limit is shrunk once by a child process
(huge_picture.py, low priority, one at a time) and cached; a card is
never a large original, one with no card image gets an SVG with its size.
- /anomalous/output_preview gives the viewers the picture or a cached
4096 px copy of a very large one; the viewer (moved to
ui_gallery_viewer.js), the image workbench's stage and the AI pages'
larger pictures use it, with a note and "Open the original".
- The listing carries each file's size (and big ones' pixel size); such a
card says "Very large".
- The browser-side PNG metadata read stops at the image data.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The panel's text for each of the 18 blocks (about, zh and en) is one short sentence; where it came from and how it was tried stay under Details, the AI's instructions (after, notes) unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 18 blocks are grouped into 7 uses (a block's "job": repaint,
new_background, upscale, refine, follow, detail, process); a block that is
not its use's default has a "want" word (best, 2x, pose, ...).
add_nodes(use=…, want=…) has reference_blocks.pick choose the first by rank
this ComfyUI can run (packs and models present, not turned off), the block
taking a second picture when one is given; block= by name still works.
begin_task's directory is a line a use with its wants, the rest by what it
needs ("LanPaint: repaint best"), so it no longer grows with the blocks and
the word budget tiers are gone. Tried live: use=new_background with two
pictures built place_in_scene.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Its CheckpointLoaderSimple takes any SD 1.5 or SDXL family checkpoint; a model in separate UNET, CLIP and VAE files (Anima) does not fit it, which the block and the changelog now say instead of "another family too".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
begin_task's text was at its 4000-character limit with every new block. The
directory now shortens in steps: every word while it fits (625 characters),
then the blocks' names with the missing ones grouped by what they need
("LanPaint: inpaint_lanpaint"), then names alone. With every pack in, the
names (about 400) leave begin_task near 3850 even with the stale tool list
line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
restyle reads a handed picture's outlines with ComfyUI's own Canny, holds
the layout with the union ControlNet (canny mode, 0.7 to 80% of the steps)
and redraws the whole picture in the prompt's new style. Tried on an RTX
4050 6 GB with SDXL (Illustrious) 832x1216: at denoise 0.8 it only got
lighter, at 1.0 it became a watercolour with the layout held, so 1.0 is
the default; about 41 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
refine_other_model redraws the made picture at denoise 0.4 with a second
checkpoint of its own (any family: its own CLIP and VAE, prompts copied from
the first, LoRAs added onto it with add_lora) and saves it beside it
(prefix refined), as the reference workflow's refiner keeps a loader and
LoRA stack apart. Tried on an RTX 4050 6 GB, Illustrious then NoobAI at
832x1216: about 66 s for the whole run, the composition kept, the style the
second model's.
A block node may be "like" a path ("$sampler.positive"): it takes the
settings of the node feeding it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
