SmartTools

Custom nodes for ComfyUI. Category: slikvik / slikvik/Image / slikvik/LLM / slikvik/Prompt.

Installation

1. Copy this folder into ComfyUI’s custom_nodes directory (or clone the repo there), e.g.
   ComfyUI/custom_nodes/SmartTools
2. Restart ComfyUI.

Dependencies match a typical ComfyUI install: PyTorch, NumPy, Pillow. Smart Save also uses ComfyUI’s folder_paths, comfy.cli_args, and the Comfy server for the save route.

Smart LLM is optional: install Transformers-related packages from requirements.txt into the same Python environment as ComfyUI (see that file for versions): C:\APPS\AI\ComfyEasyInstall\ComfyUI-Easy-Install\python_embeded\python.exe -m pip install -r C:\APPS\AI\ComfyEasyInstall\ComfyUI-Easy-Install\ComfyUI\custom_nodes\SmartTools\requirements.txt

Nodes

Smart Resizer

Display name: Smart Resizer
Category: slikvik/Image

Resizes a batch of images to a target resolution with optional letterboxing (pad) or crop to fit.

Input	Notes
image	Image batch (B, H, W, C)
use_presets	ON — target size from VIDEO_preset; OFF — from Megapixels + Multiple
resampling	Lanczos, Bilinear, or Nearest-Exact (Pillow nearest-neighbor)
Megapixels	When presets off: target size in MP (0.10–5.00)
Multiple	When presets off: both sides snapped to be divisible by this (e.g. 16)
VIDEO_preset	480p, 720p, 1080p, or 1024px (when presets on)
pad_image	On: pad (letterbox); off: crop to target aspect
Outpaint	pad_left / pad_top / pad_right / pad_bottom, feathering, optional mask, overlay_mask — applied after resize in both preset and megapixel modes

Use Presets ON: Chooses square-ish vs wide (16:9 / 9:16 style) from input aspect ratio, then applies the selected preset dimensions.

Use Presets OFF: Ignores VIDEO_preset. Chooses width/height that preserve aspect ratio, approximate total pixels Megapixels × 1_000_000, and satisfy Multiple.

Outputs: image, width, height, mask (includes outpaint / feathering when used).

---

Smart Resizer v2

Display name: Smart Resizer v2
Category: slikvik/Image

Separate from Smart Resizer (v1) for testing. Sizing is driven by Size Mode, Aspect Ratio, and Multiple (applied last).

Input	Notes
image	Image batch (B, H, W, C)
size_mode	Dimensions, Megapixels, Shortest, or Longest (shown first in the UI)
width / height	Used when Size Mode is Dimensions. 0 means derive from other settings
megapixels	Used when Size Mode is Megapixels. Target area in MP (0.10–100.00, default 1.0)
shortest / longest	Used in Shortest / Longest mode. Sets that edge after Aspect Ratio (default 1024)
aspect_ratio	Input, Smart, 1:1 (Square), 16:9 (Widescreen), 9:16 (Portrait Widescreen)
multiple	Final W/H snapped to this multiple (default 1)
pad_image	Pad (letterbox) or crop to fit
pad_colour	Letterbox colour: Black (default), Grey, Red, Green, or White
resampling	Lanczos / Bilinear / Nearest-Exact
feathering	Feathers letterbox pad edges; combined with optional mask via max
overlay_mask	Applies the selected Pad Colour where the output mask is bright
mask	Optional; resized into the content region

Dimensions + Aspect Input: both dims 0 → keep source size. Both non-zero → use that exact target canvas and pad/crop into it. Exactly one dim set → derive the other from the source aspect.

Megapixels: target area is megapixels × 1_000_000. Aspect Input preserves source aspect (v1-style). Smart / fixed ratios use that aspect while matching the area.

Shortest / Longest: set the named edge, then apply Aspect Ratio. For 16:9, Shortest is the 9-side and Longest is the 16-side. Aspect Input keeps the source aspect while resizing that edge. Smart picks 1:1 / 16:9 / 9:16 first, then applies the edge.

Smart / fixed ratio (Dimensions): both dims 0 → build a target frame of that aspect from the input (pad = contain, crop = cover). One dim set → that dim forces the frame at the target aspect. Both dims set → expand the proposed box to the aspect, then pad/crop. Multiple snaps last (keeping the locked aspect).

The node shows Input W×H, Output W×H, and a preview of the result (web/smart_resizer_v2.js).

Outputs: image, width, height, mask.

---

Smart Save

Display name: Smart Save
Category: slikvik

Same PNG output and metadata behaviour as ComfyUI’s built-in Save Image (folder_paths.get_save_image_path, %batch_num%, optional workflow metadata in PNG), but files are written to the output folder only when you click Save Image on the node.

Formats

• PNG / WebP / AVIF: embed prompt + workflow so drag-and-drop onto ComfyUI restores the graph (WebP/AVIF use the same EXIF prompt: / workflow: tags ComfyUI reads).
• JPEG: EXIF stores the CLIP prompt (ImageDescription / UserComment) and seed (Artist / UserComment). JPEG cannot restore a workflow on drop.

1. Queue the workflow at least once so the node can cache the current image batch (keyed by the graph node id).
2. The node shows a temp preview so you can review the image without writing to the final output path.
3. Click Save Image to write the selected format to the output directory.

Requires the included web extension (web/smart_save.js); restart ComfyUI after installing or updating this pack.

---

Smart Save Video

Display name: Smart Save Video
Category: slikvik

Self-contained FFmpeg video encoder with the same manual-save workflow as Smart Save. It does not depend on VideoHelperSuite. Bundled presets live in video_formats/ (h264/h265 MP4, WebM, AV1, NVENC variants, ProRes, FFV1).

Input	Notes
images	Frame batch (B, H, W, C) to encode.
autosave	Default OFF. When ON, the final video is written during execution; the Save Video button is disabled.
inc_audio	Default ON. When ON and audio is connected, audio is muxed into the single saved file. When OFF (or no audio), exactly one silent video is produced. Never writes both silent and muxed copies.
frame_rate, loop_count, pingpong	Encoding timing controls.
filename_prefix	Relative subfolder/name, plain prefix, or absolute folder path.
format	Bundled FFmpeg preset name. Default h264-mp4.
crf, pix_fmt, bitrate, megabit, profile	Quality / codec options applied when the selected preset uses them.
save_metadata	Embed workflow/prompt metadata into the container when the preset supports it. No first-frame PNG sidecar is written.
audio	Optional ComfyUI AUDIO dict (waveform, sample_rate).
original	Optional frame batch. When connected, the node encodes a second silent temp preview and the UI shows New and Original side-by-side, time-synced from the first frame. Save / autosave still target only the primary (images) video.
use_timestamp	Counter vs timestamp filenames on final save.

ffmpeg: Required for encoding. Resolved from SMART_SAVE_VIDEO_FFMPEG / VHS_FORCE_FFMPEG_PATH, imageio-ffmpeg, system PATH, or a local ffmpeg / ffmpeg.exe.

1. Queue the workflow so the node encodes a temp preview video. The node UI shows an interactive video player (play / pause / mute / timeline). When original is connected, both videos play side-by-side and stay synced.
2. With autosave OFF, click Save Video to copy that encoded file to the chosen output path without re-encoding.
3. Buttons match Smart Save: browse path, open folder, favorites, and save history (web/smart_save_video.js).

Outputs: video_path is the absolute path of the autosaved file, or the temp encode when autosave is off.

---

Smart Load Video

Display name: Smart Load Video
Category: slikvik

Self-contained FFmpeg video loader. It does not depend on VideoHelperSuite. Soft-modeled on VHS Load Video FFmpeg (Upload) for the core load widgets.

Input	Notes
video	Combo of videos already in the ComfyUI input folder. Drag-and-drop can add a small file here; ComfyUI rejects uploads larger than --max-upload-size (default 100 MB).
video_path	Absolute path (or ~) to a video on disk. When set, this is used instead of the combo — FFmpeg reads the file in place, so size does not matter. Use browse video path.
force_rate	0 keeps the source fps. The widget shows the source rate (59.93↩) when at 0. Disable (slashed circle) returns to 0; reset writes the source fps as an explicit value.
custom_width / custom_height	0 keeps or derives that edge from the other. Both set crops to the new aspect, then scales. Format presets add reset sizes (H3: 1344×768).
multiple	After the target size is chosen, both edges snap up to this multiple. Default 1 is a no-op. Selecting a format other than None sets this to the preset grid (H3 → 32). You can override it afterward.
rotation	Clockwise last-step turn of the loaded IMAGE and mask (0, 90, 180, 270). 90 / 270 swap H×W after custom size. The in-node preview uses the same angle.
format	VHS-style model preset (None, AnimateDiff, Mochi, LTXV, Hunyuan, Cosmos, Wan, H3). Only updates reset/step targets until you click reset. H3: reset 24 fps and 1344×768, multiple 32. Requested caps snap up to 5 + 17n (48 → 56); leftover loaded frames still truncate down.
cap_seconds	0 leaves frame_load_cap alone. Otherwise live-sets frame_load_cap from loaded fps × seconds (force_rate if set, else source fps), then snaps up to the format frame rule (H3: 2s at 24 fps → 56).
frame_load_cap	0 loads all remaining frames from the effective start. The widget shows the source (or format-legal) frame count when at 0. Disable returns to 0; reset writes that count. Overwritten while cap_seconds is > 0.
start_time	Start offset in seconds (index 0).
slice_index	0 starts at start_time. Each next index jumps forward by frame_load_cap frames (start_time + slice_index * frame_load_cap / fps). slice_index > 0 requires frame_load_cap > 0.

Large clips: use video_path (browse or paste), leave start_time at 0, set cap_seconds (e.g. 10) so frame_load_cap is filled and format-snapped, and increment slice_index each run (a 2-minute clip in 10-second chunks is indices 0–11).

The node shows a preview of the selected combo file or disk path, clipped to the loaded slice (start_time + slice_index * frame_load_cap / fps, for frame_load_cap / fps seconds). It annotates force_rate / frame_load_cap with the source fps and total frames (web/smart_load_video.js). Restart ComfyUI after installing or updating this pack so the browse/view/query routes register.

Outputs: IMAGE (frame batch), mask (inverted alpha, or ones when there is no alpha), audio (Comfy AUDIO aligned to the loaded slice; silence if the file has no audio), framerate (loaded fps: force_rate when set, otherwise source fps), total_frames (frames actually loaded after format trim).

ffmpeg: Same resolve order as Smart Save Video (SMART_SAVE_VIDEO_FFMPEG / VHS_FORCE_FFMPEG_PATH, imageio-ffmpeg, system PATH, or a local ffmpeg / ffmpeg.exe).

---

Smart Lora

Display name: Smart Lora
Category: slikvik

Applies two independent lists of model-only LoRAs in one node: high LoRAs to the high-noise model and low LoRAs to the low-noise model (e.g. for split high/low-noise model setups). No CLIP is touched.

Input / Output	Notes
model_high (in/out)	Optional. High-noise diffusion model; high LoRAs are applied to it. If left unconnected, the output is None.
model_low (in/out)	Optional. Low-noise diffusion model; low LoRAs are applied to it. If left unconnected, the output is None.
prompt (in)	Optional string from another node; if connected it is prepended to the prompt text with a line break.
prompt (out)	Combined prompt: optional prompt input, line break, then prompt_text.

LoRA lists (custom UI): Use Add Lora (High) / Add Lora (Low) to add rows to each group. Each row has:

• a name field (click to open a searchable LoRA picker),
• a strength box (click to type a value; negative values allowed),
• an info button (i) that, when a sidecar JSON exists next to the LoRA file (same name, .json extension), opens a modal showing the link, trigger words, and description. Description may be a string or { "model": "...", "version": "..." }; both HTML fields are shown in the same box, and basic tags such as <p>, <ul>, <li> are rendered. Each field is copyable to the clipboard,
• an enable toggle, and
• a delete button (✕).

High LoRAs apply to model_high and low LoRAs to model_low, in list order, only when their toggle is on and strength is non-zero.

Profiles: Above the prompt text box, a Profile selector with Save Profile As..., Update Profile, and Delete Profile buttons lets you store and recall setups. A profile captures both LoRA lists (each LoRA's name, strength and enable toggle) plus the prompt text. Selecting a profile loads it (replacing the current lists and prompt text), Update overwrites the selected profile with the current state, and Delete removes it. Profiles are stored globally on the server in smart_lora_profiles.json, so they are shared across every Smart Lora node and all workflows, and persist across restarts.

Resizing: Drag the node wider/narrower and the rows reflow horizontally. Drag it taller/shorter and only the prompt_text box grows or shrinks; the LoRA rows and buttons stay fixed.

Persistence: The full LoRA configuration is stored as JSON on the node (in node.properties / the hidden lora_config input) and is sent to the backend, so workflows reload exactly as saved.

Requires the included web extension (web/smart_lora.js); restart ComfyUI after installing or updating this pack.

---

Smart LLM

Display name: Smart LLM
Category: slikvik/LLM

Runs local Hugging Face vision-language instruction checkpoints from a model folder (full snapshot: config.json, tokenizer / processor files, and *.safetensors or a sharded model.safetensors.index.json). Inference uses Transformers (AutoModelForImageTextToText + AutoProcessor); there is no GGUF or llama.cpp dependency.

Works with any checkpoint those Auto classes load — for example Qwen3-VL (including fine-tunes such as Huihui abliterated builds) when AutoProcessor succeeds, and Gemma 4 (with a Gemma-specific processor fallback if AutoProcessor fails). There is no separate “model family” toggle.

Input	Notes
model_folder	Absolute or ~ path to the directory you downloaded (e.g. with huggingface-cli download ... --local-dir ...). Must contain config.json and safetensors weights.
system_prompt	Optional multiline system message.
prompt	Multiline user prompt.
max_tokens	Cap on new tokens decoded after the prompt (same idea as max_new_tokens in Transformers). Increase if output looks truncated.
attn_implementation	Transformers attention: sdpa (default), eager, or flash_attention_2 (requires flash-attn + CUDA; falls back to SDPA with a warning if unavailable).
device_placement	cuda (default) frees Comfy-cached models and forces the complete HF model onto GPU, avoiding unpredictable CPU offload. auto lets Accelerate split/offload the model when VRAM is insufficient, which can make token generation dramatically slower.
enable_thinking	Enables reasoning in chat templates that support it. Default OFF for direct responses; reasoning can consume the full token budget before the final answer.
unload_model	ON — after generation, drop the model and processor from memory and call CUDA cache cleanup so later nodes get more VRAM. OFF — keep the model loaded for the next run (same model_folder path).
video_fps	Frame rate of the optional video batch (default 30). Match VideoHelperSuite force_rate / loaded fps so temporal grounding is correct. Ignored when video is disconnected.
max_video_frames	Cap on frames taken from video (even subsampling). Default 32. 0 = use all frames, but batches larger than 64 are auto-capped (avoids multi‑minute hangs / VRAM blowups). Prefer VHS frame_load_cap for long clips. Ignored when video is disconnected.
image	Optional. First batch frame as RGB PIL inside the HF processor.
image_2	Optional. Second batch frame, after image, passed as PIL to the processor.
video	Optional. Video as an IMAGE frame batch (B, H, W, C) — same type as VideoHelperSuite Load Video IMAGE output. All frames are sent as native HF video (not as still images).

Sage Attention vs Smart LLM: ComfyUI’s Sage Attention (the sageattention package and flags such as --use-sage-attention) plugs into diffusion sampling (comfy attention). Smart LLM does not use Sage; the VL model runs inside Hugging Face and only supports the backends above—not sageattention.

VRAM: Full VL checkpoints are large; use a size and dtype your GPU can hold, or explore quantization in Transformers separately. Video adds frame tokens — prefer VHS frame_load_cap and/or max_video_frames if generation OOMs. CUDA uses device_map="auto" (requires accelerate). With Unload OFF, the same model_folder + attn_implementation pair reuses one in-memory model (no reload each run). Changing model_folder drops any cached weights for other folders first. Changing attention backend replaces the previous cached copy for that folder so VRAM does not stack. (If you use several Smart LLM nodes with different folders in one workflow, the cache holds one folder at a time—whichever loads last—so the other path may reload on its next run.)

Transformers version: See requirements.txt. Gemma 4 needs transformers 5.5.0 or newer; Qwen3-VL needs a Transformers build that registers its processor/model classes.

Smart H3 LLM

Display name: Smart H3 LLM

Category: slikvik/LLM

Uses the same local Hugging Face model cache as Smart LLM to analyze references and produce a duration-aware prompt that follows the bundled MiniMax H3 prompt-writing guides. It runs a factual media-analysis pass, generates the H3 prompt deterministically, validates its structure and timing, and makes up to two text-only repair attempts if needed.

Input	Notes
skill	base for T2VA/I2VA/FL2VA/L2VA, or ref2VA for full-reference six-section output.
base_workflow	Explicit base workflow. Ignored by ref2VA.
ref_image_1_role…ref_image_4_role	ref2VA only: assign each image independently as Auto, subject/reference, first frame, intermediate keyframe, last frame, or storyboard/composition. A disconnected image's role is ignored.
ref_video_role	ref2VA only: infer from the prompt, use the video for reference generation, treat it as a video editing source, or perform video continuation.
prompt	Creative intent and desired use of references. The base workflow does not need to be repeated here.
verbatim_dialogue	Optional dialogue/lyrics to preserve exactly inside H3 <d> blocks.
video_duration	Exact target duration in seconds (default 15.00); controls final-frame alignment and valid cut range.
visual_style	Auto or a specific H3-compatible visual style hint.
shot_count	0 lets the model choose; a positive value is validated as an exact shot count.
audio_usage	Infer from prompt, copy/reuse, reference only, or ignore connected audio.
max_tokens	Generation budget, default 4096 to accommodate detailed ref2VA output.
device_placement	cuda by default to prevent intermittent CPU-offloaded runs; use auto only when the checkpoint cannot fit fully in VRAM.
enable_thinking	Default OFF and normally should remain off for H3. Thinking can exhaust max_tokens before required fields such as detailed_description.
image_1…image_4	Optional still references. Base workflows use only their prescribed sockets; ref2VA numbers connected images densely.
video	Optional VHS-style IMAGE frame batch, with video_fps and max_video_frames matching Smart LLM.
audio	Optional standard AUDIO dictionary, directly connectable from VideoHelperSuite Load Audio.

Base image mapping:

• T2VA: text-driven; connected still images are not used or labeled.
• I2VA: requires image_1, mapped to <Picture 1> at 0.00 seconds.
• FL2VA: requires image_1 as Picture 1 at 0.00 and image_2 as Picture 2 at the exact final duration.
• L2VA: requires image_1, mapped to <Picture 1> at the exact final duration.

ref2VA uses all connected images in socket order, skipping gaps, then exposes the video and audio as <Video 1> and <Audio 1> when active. Each connected image keeps the role selected for its original socket: if only image_1 and image_3 are connected, they become <Picture 1> and <Picture 2> while using ref_image_1_role and ref_image_3_role. Reusable people, objects, scenes, styles, or actions are defined separately as <Subject N> by the model.

base_workflow remains separate from ref2VA task types because base workflows are mutually exclusive, while a ref2VA summary may combine several relationships. First/intermediate/last-frame roles require keyframe completion; subject/reference and storyboard/composition roles require reference generation. These can combine with the selected video and audio tasks—for example, a character-reference image + a last-frame image + an edited source video + reference-only audio requires [keyframe completion + reference generation + video editing + audio reference]. Any role left on Auto from prompt is inferred from the user prompt and media analysis.

The audio socket accepts {"waveform": Tensor[B,C,T], "sample_rate": int}, the same payload returned by VideoHelperSuite LoadAudio. Audio is mixed to mono and resampled to 16 kHz for inference. The selected checkpoint must support every connected modality; Gemma 4 E2B, E4B, and 12B variants support native audio, while many vision-language checkpoints do not. If source-audio and target-video durations differ, the prompt is instructed to describe partial reuse, trimming, continuation, or padding rather than claiming impossible 1:1 reuse.

Outputs: h3_prompt is a clean paste-ready H3 prompt with no analysis or markdown wrapper. analysis contains the model's factual notes about the connected references and can be left unconnected.

Because this node normally performs two model generations (analysis and prompt writing), with up to two additional generations when repair is required, it takes longer than a single Smart LLM call. The model remains loaded between passes and follows the unload_model setting after completion.

Smart H3 Prompt

Display name: Smart H3 Prompt

Category: slikvik/Prompt

Builds a paste-ready MiniMax H3 prompt entirely with deterministic Python logic. It does not load an LLM or inspect media. Instead, describe the future reference assets in the multiline <Picture 1>…<Picture 4>, <Video 1>, and <Audio 1> widgets.

Input	Notes
skill / base_workflow	Select base T2VA/I2VA/FL2VA/L2VA output or the six-section ref2VA format.
Reference-role selectors	Assign each described picture and video its H3 role. In ref2VA, Auto from prompt deterministically falls back to subject/reference for pictures and reference generation for video.
Shot 1	Required first-shot description. Do not type a [Shot 1] header.
Shot 2, Shot 3	Optional later-shot descriptions. Shot 3 requires Shot 2. The node adds sequential H3 headers.
Shot start times	Numeric seconds for Shots 2 and 3. The node formats them as MM:SS.mmm; populated times must increase and remain before video_duration.
verbatim_dialogue	Exact dialogue/lyrics to verify. Place each line in the appropriate shot inside <d>[Language] ...</d>; the node cannot infer its speaker or timing.
H3 section details	Optional subject_definitions, summary, retention_analysis, overall_soundscape, and non_diegetic_music bodies. Omit field headings. Missing content receives conservative generic defaults.

Base workflows require the prescribed picture descriptions: none for T2VA, Picture 1 for I2VA/L2VA, and Pictures 1–2 for FL2VA. Base output keeps video/audio descriptions as ordinary prose because <Video N>, <Audio N>, and <Subject N> labels belong only to ref2VA.

ref2VA picture descriptions must be populated contiguously from Picture 1. The node creates task prefixes from the selected roles, definitions for omitted active labels, and conservative retention markers. Custom section text remains authoritative, but it must use consistent H3 labels. Generic defaults provide valid structure; for production-quality detail, explicitly fill the section fields and all shot descriptions.

The output is rejected with a focused error if references are missing, cuts are invalid, dialogue tags are malformed, or the assembled result violates the bundled H3 guides. There is no LLM repair pass.

Author

slivik (Smart Resizer header: v1.0.0 initial release).