FVM · Dataset Prompt List

You want a character LoRA dataset. The advice is consistent and completely unhelpful in practice: 15–25 images, mixed close-ups and full body, different angles, different lighting, different clothes, diverse backgrounds. Then you sit down to actually write twenty distinct prompts by hand and produce four of them before boredom wins.

Dataset Prompt List is the workaround. One prompt per line, an optional [close] / [half] / [full] shot tag in front, wildcards for everything you want re-rolled, and a variations multiplier. The shipped starter text is 20 lines - 8 close, 6 half, 6 full - and with variations = 5 that's 100 images off one text box.

Inputs

• prompt_text - the list. # lines are comments, a leading [close], [half] or [full] marks the shot size.
• shot_filter - all, close, half or full. Filtering to close and generating a batch of head-and-shoulders shots, then switching to full, is a much more controllable way to build a set than trusting a mixed list to balance itself.
• variations (1–200) - how often each line is used. Every copy re-rolls the wildcards - clothes, background, light - while the pose and angle of that line stay put. That's exactly the property you want: pose diversity is deliberate, everything else is randomised.
• order - rounds (all lines once, then all again - safe to cancel early, you still have every angle), per_prompt (all variations of line 1, then line 2…), or shuffle.
• start_index and max_rows - a window into the list, counted after the shot filter, so you can generate 40 images tonight and the next 40 tomorrow without redoing the first lot. The Preview button shows the numbering.
• seed - one seed for all wildcard draws. Same seed plus same text plus same wildcard files gives byte-identical prompts, which is how you re-render a set you liked.
• prefix / suffix - prefix takes your trigger word or "the woman from the reference image"; suffix defaults to __dataset/photo__, a realistic camera description. Wildcards work in both, and variables bound in prefix with ^var can be recalled in the lines as __^var__.

Outputs

prompt is a list - prefix, line and suffix resolved - and downstream nodes run once per entry, the same way CR Prompt List behaves. caption is the same list without prefix or suffix: that's your training caption. Keeping the two apart is the smart part of this node. The caption describes the clothes and the background so they stay adjustable instead of getting welded to the character, while the image prompt still carries the style and quality terms you don't want in the caption. shot gives you close/half/full per prompt - ideal for subfolders or filenames - index gives line numbers, count the total, and listing a readable overview with counts per shot size.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

Restart ComfyUI; ComfyUI Manager users can search comfyui-fvmtools. The prompt and JB tools need nothing beyond numpy and pydantic, both already in the ComfyUI venv. Wildcard files live in the pack's own wildcards/ folder and the toolbar's Wildcards button edits them; the syntax is __name__, __cat/*__, {a|b}, {2$$a|b|c}.

Traps

The list output is the thing people trip over. A list into a normal input runs the graph once per entry - which is the feature - but if you accidentally feed caption somewhere that expects a single string, you get one run per caption and a folder of files you didn't ask for. Save the sidecars from caption, generate from prompt, and don't swap them.

Wildcard files are read at queue time, so editing them changes your next run and silently changes what a "reproducible" seed produces. If a set matters, copy the wildcard files next to the project.

Shot tags are free text you type, not a validated enum - [Close] with a capital C is an untagged line that only shows up under all.