- An override family could declare a step such as 3; sizes rounded to it
are not divisible by 8, so create_empty_latent_batch raised ValueError
at execution instead of producing a latent (Codex review on PR #108).
- validate_table now rejects a step that is not a multiple of 8 with a
message naming the latent grid; the shipped steps (16, 32, 64) already
comply, and the schema docs state the rule.
- table validation accepts sides 64..8192, matching the widget min/max
- parse_size_label (Python and JS) caps digits and rejects out-of-range
sizes so create() falls back to width/height instead of overflowing
- family_labels() and the combo default now read the merged table, so an
appended family is selectable and a removed one is not offered
- an alias used as the override default is normalised to its label
- the frontend rebuilds the model options from the payload and keeps a
saved family that is no longer listed so old workflows still show it
- Set resolution to custom whenever the panel has no selection (wired,
failed, non-finite or loading) so a stale label cannot override width/height.
- Compute label megapixels with 1 MP = 1,048,576 px, matching the panel
header, with half-up rounding so Python and JS labels agree on ties.
- Add a native resolution COMBO (shipped sizes only) so a size can be picked
and saved in the workflow without the panel; custom or unparseable labels
fall back to width/height on both sides.
- Mirror the dropdown with the chips: options follow the active chip, the
value tracks the current size, and picking a label writes width/height.
- Draw a decorative ratio shape on each ratio chip and show the first
families warning in the status line.
- catch RecursionError alongside ValueError when loading, merging and validating
- add tests for deep nesting, remove vs families, and a default naming a removed family
- add override.py: override_path, merge_override, mtime-cached current_table
- resolve_family and families_payload consult the merged table and expose warnings
- an invalid override is ignored whole with a logged warning
- treat ints as finite instead of calling math.isfinite on them
- bound step (4096), megapixels (64) and side length (65536) with normal problem messages
- cover step, megapixels, default and size values of 10**400
- check default, labels and aliases are strings before set membership
- reject non-finite step and megapixels via a shared _is_step helper
- cover unhashable and Infinity/NaN inputs in the garbage test
- ship the table as families.json with per-family notes from the old source comments
- add table.py: load_table, pure validate_table and build_families
- families.py loads and validates at import, raising RuntimeError on problems
- include *.json as package data
- VALIDATE_INPUTS lets unknown family labels reach create()
- no size is listed under two chips; SD 1.5 3:4 gets true 3:4 sizes
- model change before the families payload loads is applied on load
- custom ratio: clear error on chip click, cap sides at 8192, accept X
- real-host assertions for model change and wired-in state
- add ModelResolutionPickerNode rounding to the family step
- serve GET /kikotools/model_resolution/families
- register the node in NODE_CLASS_MAPPINGS
- add CHIPS, MIRROR, Family and the FAMILIES literal table
- add family_labels, resolve_family and families_payload helpers
- cover step, ratio, mirror and sort invariants with tests
ComfyUI resizes an empty latent to the model's latent format only when the
latent carries downscale_ratio_spacial, as core EmptyLatentImage sets it.
Without it, models with 16x VAEs (Flux.2, Klein, MOSS-Image, Qwen-Image 2.1)
sampled at twice the requested width and height.
- Split run_name on "/" and "\" and reject empty, "." and ".." parts, so
a Windows-style "..\foo" can no longer resolve outside model_compare
and let the index-0 archive step rename an unrelated output folder
- Reject drive-letter names and require the resolved directory to sit
strictly below output/model_compare after symlinks are followed
- Nested run names such as project/date/sweep still work
- Checkpoint Folder Iterator loads one checkpoint per queued run from a
folder (relative to models/checkpoints or absolute), so a whole folder
can be swept with only one model in memory at a time
- Model Compare Sheet saves each run's image as a tile under
output/model_compare/<run_name>/ and rebuilds a labeled contact sheet
from disk every run, so the grid survives between independent prompts
- Starting a sweep at index 0 archives stale tiles to a timestamped folder
- Running past the last checkpoint raises a clear error instead of
wrapping, so an over-queued batch never restarts and archives a
finished sweep
- Checkpoints load from the scanned absolute path, so folders outside the
configured roots work as the tooltip promises
- Extract the LoRA Folder Batch folder-scanning, natural-sort and regex
filter helpers into kikotools/base/model_folders.py so both tools share
one implementation
- google-generativeai is deprecated and emits a FutureWarning at node
load; google-genai is Google's supported replacement
- Image analysis now uses genai.Client().models.generate_content
instead of the module-level configure/GenerativeModel API
- Model listing reads supported_actions, the new SDK's name for
supported_generation_methods
- Install instructions, requirements.txt, and README updated to the
new package name
Drop RETURN_NAMES so ComfyUI shows the output as STRING, matching the original Project File Path Generator. Links connect by slot index, so existing workflows are unaffected.
Builds a save prefix like MyProject/2026-10-13/images/image for a save
node's filename_prefix, replacing djz-nodes' Project File Path Generator
with a fresh implementation:
- Same inputs, root defaults to empty (prefixes are already relative to
the output folder, so "output" nested as output/output/...)
- Optional date folder; IS_CHANGED tracks the date instead of forcing a
re-run on every queue
- Components are split on / and \ with '..' segments dropped, so a path
can't climb out of the output folder
- Only real file extensions are stripped (render_v1.5 keeps its dot)
- The separator option is applied last instead of being undone by
normpath; an empty filename raises a clear error
- Close PIL images and NamedTemporaryFile handles before deleting the
files; Windows refuses to delete open files (WinError 32)
- Set USERPROFILE alongside HOME: expanduser ignores HOME on Windows
- Compare realpaths in the nested save-prefix test; Windows temp dirs
use 8.3 short names (RUNNER~1)
save_paths is reachable without auth, so saving "/" made every media
file on the machine servable. Reject filesystem/drive roots, the home
directory and its ancestors (realpath-resolved, so symlinks count), and
ignore such entries already in config.json.
ComfyUI registers custom_nodes in folder_names_and_paths, so treating
every registered folder as a download root let a queued prompt drop a
.py file there that runs on the next restart.
- Skip custom_nodes when collecting allowed roots
- save_response only writes model/data extensions (.safetensors, .gguf,
.ckpt, .pth, .json, ...), so any future code-holding folder is still
protected
- KikoSaveImage: reject absolute or '..' filename_prefix directories that
resolved outside the output folder
- Film grain: use a local torch.Generator instead of reseeding the global
RNG; strength 0 no longer skips the toe adjustment
- Purge VRAM: IS_CHANGED returns NaN so caching never skips the purge
- Resolution calculator: clamp outputs to >= 8
- Plot sampler params: font size floor of 1 for images under 32px wide
- Gemini prompt: drop Image.fromarray mode= argument (removed in Pillow)
- Reject filenames with path components; reduce server-derived names
(Content-Disposition, CivitAI redirect, URL) to a safe basename
- save_path must resolve inside models_dir or a registered model folder
(including extra_model_paths.yaml roots); relative paths resolve
against ComfyUI's base path
- Only http/https for user URLs and redirects; drop the Authorization
header on cross-host redirects
- Stream to <file>.part and rename on success, removing the partial file
on any failure, so a dropped connection no longer leaves a truncated
model that later runs treat as complete
- 60s timeout on every request; shared streaming loop in BaseDownloader
- Custom downloader honours Content-Disposition before the default name
/view served any readable file (e.g. ~/.ssh/id_rsa) to anyone who could
reach the server; the only guard was a '..' substring check.
- Resolve paths with realpath and require commonpath containment in the
saved paths or ComfyUI input/output directories
- /view serves image/video/audio only, /thumbnail and load_media images
only; the extension check applies to the resolved path, so symlinks
can't smuggle other files
- /images 403s outside the roots; page/per_page validated and clamped
- /list_directories (folder picker) returns directory names only
- load_media refuses disallowed paths from a tampered selections.json
- Route handlers moved to module level so they are unit tested
- Update stale tests to the (image, info) return contract
- LoRA strength strings: step <= 0 looped forever and reversed ranges
returned [] (ZeroDivisionError later); now raise ValueError, and cap
generated values at 1000
- Flux sampler params: unparseable steps/guidance/shift/denoise/seed
tokens raise with the field name instead of being dropped, and an
empty LoRA folder raises instead of returning (None, [])
- Detect Flux schnell via unet_config guidance_embed (the ModelType.FLOW
check never matched Flux)
- IS_CHANGED: re-run Flux params when a seed is '?'; LoRA folder batch
hashes file list + mtimes instead of always re-executing
- LoRA folder scan uses commonpath containment across all lora roots
- Prompt separators must be whole lines; failed encodes raise
Previously split_latent_batch and join_latent_batch only handled the
"samples" key, silently dropping metadata like noise_mask or batch_index.
This broke inpaint and masked workflows after a round-trip conversion.
Add 4 utility nodes for converting between batched tensors and lists,
eliminating the dependency on Impact Pack for this common operation:
- ImageBatchToImageList (OUTPUT_IS_LIST)
- ImageListToImageBatch (INPUT_IS_LIST)
- LatentBatchToLatentList (OUTPUT_IS_LIST)
- LatentListToLatentBatch (INPUT_IS_LIST)
JS Math.random() only has 53 bits of integer precision, making the
64-bit range produce non-uniform values. 32-bit (0xFFFFFFFF) is the
standard ComfyUI seed range and works correctly in both Python and JS.
Also removes the redundant mode input from SeedHistoryNode, relying on
ComfyUI's built-in control_after_generate widget instead.
Refs #43
ComfyUI is converting nodes_latent.py to V3 Schema on October 8th, which will
break direct imports of LatentBatch. This commit replaces the import with a
local implementation copied directly from ComfyUI source code.
Changes:
- Removed: from comfy_extras.nodes_latent import LatentBatch
- Added: Local batch_latents() and reshape_latent_to() functions
- Updated: latentbatch.batch() calls to use batch_latents()
- Added: torch and comfy.utils imports for tensor operations
- Added: Comprehensive unit tests for latent batching functionality
The local implementation is functionally identical to the original and ensures
the node will continue working after the V3 schema migration.
Test Coverage:
- 5 new tests in TestLatentBatchingFunctions class
- All 16 tests passing (11 existing + 5 new)
- Tests cover tensor operations, batch indexing, and reshape logic
- Add auto-batching functionality to split large LoRA collections into manageable chunks
- Implement batch_size parameter to control number of LoRAs per batch (default: 25)
- Add batch_index parameter to select which batch to process
- Include batch tracking metadata in LORA_PARAMS for visualization
- Display batch info in plot parameters when available
- Update lora_list output to show batch header when auto-batching is enabled
- Add comprehensive tests for batching functionality
- Update documentation with detailed auto-batching usage instructions
This feature prevents UI disconnection issues when processing large numbers of LoRAs
by allowing users to process them in smaller batches sequentially.
Vito Sansevero