Fal lora inference

fal LoRA Inference

Generate images from trained LoRA weights with fal-ai/flux-lora, save them locally, and judge whether the LoRA is usable.

Defaults

Setting	Default
Endpoint	fal-ai/flux-lora
LoRA scale	1.0
Image size	portrait_4_3
Inference steps	28
Guidance scale	3.5
Images per prompt	1
Output format	jpeg
Trigger in prompt	Required

Paid-submission rule

Generation is a paid external action.

• If the current user request explicitly says to generate, render, run inference, sample, or test the LoRA, proceed without asking twice.
• Otherwise, show the endpoint, prompts, and settings and ask for confirmation.
• Pass --yes only after one of those conditions is met.

Cost scales with prompts × scales × --num-images. State the total image count before a large sweep.

The trigger phrase is not optional

The trigger phrase carries the identity. A prompt without it renders a generic person even though the LoRA loaded successfully, which reads as a broken LoRA. The script refuses such prompts unless --no-require-trigger is passed.

Workflow

Resolve the project-local skill first:

SKILL_DIR=".cursor/skills/fal-lora-inference"
[ -d "$SKILL_DIR" ] || SKILL_DIR="$HOME/.cursor/skills/fal-lora-inference"
[ -d "$SKILL_DIR" ] || SKILL_DIR="skills/fal-lora-inference"

1. Confirm inputs

Ask if missing:

• LoRA weights: a fal URL, or a local .safetensors path to upload
• Trigger phrase used at training time
• Prompts, or the intent to sweep with defaults
• Output directory

The trigger phrase and weights URL are recorded in training_result.json and training_request.json if fal-lora-training produced this LoRA. Read them instead of asking.

2. Sweep the LoRA scale

For an untested LoRA, start with one prompt across several scales at a fixed seed so composition stays constant and scale is the only variable:

.venv/bin/python "$SKILL_DIR/scripts/generate_lora_images.py" \
  --lora "./portrait_training_run/subject_lora.safetensors" \
  --trigger-phrase "ohwx man" \
  --prompt "ohwx man, close-up portrait, soft window light, navy blazer" \
  --lora-scale 0.7 0.9 1.1 1.3 \
  --seed 12345 \
  --output-dir "./lora_generations/scale_sweep" \
  --env ".env" \
  --yes

3. Vary the prompts

Once a working scale is known, hold it and test range — shot distance, lighting, wardrobe, setting, expression:

.venv/bin/python "$SKILL_DIR/scripts/generate_lora_images.py" \
  --lora "https://…/pytorch_lora_weights.safetensors" \
  --trigger-phrase "ohwx man" \
  --prompts-file "./prompts.txt" \
  --lora-scale 1.0 \
  --image-size portrait_16_9 \
  --num-images 2 \
  --output-dir "./lora_generations/prompt_set" \
  --env ".env" \
  --yes

--prompts-file takes one prompt per line and ignores blanks and # comments.

The script uploads a local LoRA if needed, runs prompts × scales concurrently, downloads every image, writes generations.json, and renders contact_sheet.jpg.

4. Review the output visually (required)

Open contact_sheet.jpg with image vision, then open individual files at full resolution. Do not report success from exit codes alone — a completed request says nothing about likeness.

Judge likeness, skin texture, eye consistency, prompt adherence, and pose variety against references/evaluation-checklist.md.

5. Deliver

Report:

• Working LoRA scale, and what failed outside it
• Whether the LoRA looks under- or overtrained, with the evidence
• Output directory and contact sheet path
• Seeds worth reusing
• Any failed requests

Recommend a retrain only with a specific reason, such as likeness needing 1.3, or training wardrobe persisting across every prompt.

Authentication

Requires fal-client, python-dotenv, and either FAL_KEY or FAL_API_KEY. The script maps FAL_API_KEY to the client-compatible FAL_KEY without printing it. pillow is optional and only powers the contact sheet.

Scripts

Script	Purpose
scripts/generate_lora_images.py	Upload LoRA, run prompt × scale matrix, download images, write manifest and contact sheet

Degrees of freedom

• Low: trigger phrase present in every prompt, local downloads, visual review before reporting
• Medium: prompt wording, image size, steps, guidance, seeds
• High: which scales to sweep, and whether results justify a retrain

See references/flux-lora-schema.md for the full input and output schema.