Photo Editing (FLUX.2)

Use this skill for every request to modify an uploaded real photo: retouching, restyling, photo-to-photo generation, Fuji/底片/膠卷 looks, color grading, bokeh, or content changes.

The deployed entry point is already available. Never write a new Diffusers/SD script and never download models.

cd /home/chihmin/models-work/flux2
source .venv-rocm72/bin/activate
FLUX2_BIG_WMMA_LINEAR=1 python ~/.hermes/skills/create-image/scripts/create_image.py "PROMPT" --image /path/to/photo.jpg

The command returns JSON. Attach the final_path PNG in the same Discord reply:

[[image: /absolute/path/to/result.png]]

Required workflow

1. Open and inspect the input image first. Identify its source type, orientation, subject, composition, lighting, and what must remain unchanged.
2. Research first when the request names an unfamiliar style, person, location, franchise, product, or other meaningful term.
3. Choose the route below.
4. Attach the resulting image; do not reply with only its filesystem path.

Route selection

Source/request	Route
Real photo; change only photoreal look while keeping composition	--refcontrol --image
Real photo; Fuji, analog, film, 底片, 膠卷, F1.2 film look	--image + film wording (auto RefControl film flow)
Real photo; freeform content edit	--image
Photo → anime / illustration / PVC	Do not use this skill’s FLUX path; use create-image with --anime --image
Source is an anime illustration or sketch → anime restyle/CG	Use create-image with --anime --image

An uploaded image must be passed with --image; never ignore it and run text-to-image unless the user explicitly says it is only background context.

Photoreal style changes: RefControl

For a real photo whose style should change while its scene/layout should remain, use --refcontrol. It uses the source photo plus a Depth Anything V2 depth map and a RefControl depth LoRA to structure-lock the edit.

python ~/.hermes/skills/create-image/scripts/create_image.py \
  "cool Fuji green-cyan tones" \
  --refcontrol --image /path/to/photo.jpg

Prompt discipline

• Style-only request: use only short target-style words. Do not repeat the subject, pose, scene, objects, composition, or generic quality tags; the photo already contains them and verbose content descriptions cause redraws.
  • Good: cool Fuji green-cyan tones, Leica reportage, F1.2 shallow depth of field
  • Bad: a long description of the person/room/pose already in the photo.
• Content edit: state what must be retained and what changes, e.g. Keep the person, pose and composition; change the background into a rainy neon Tokyo street.

RefControl controls

• --refcontrol-scale 1.0: structure adherence; leave at default unless the user requests more/less layout preservation.
• --lora-scale: style/film strength. For regular film RefControl the default is 0.55; 0.4–0.5 is subtle, 0.8–0.9 strong.
• Keep --seed fixed while adjusting a single parameter.
• RefControl is 9B-only. Do not use --fast-preview with it.
• Typical 2-step portrait run: about 54 seconds end-to-end. Default is 5 steps unless specified.

Fuji / analog / film / 底片 edits

Film words plus --image automatically select the structure-locked RefControl film route and load Analog Redmond. Do not add --refcontrol manually unless useful for clarity.

python ~/.hermes/skills/create-image/scripts/create_image.py \
  "底片風" --image /path/to/photo.jpg

For the request 「F1.2 底片風」 use this exact style base (append only extra user-requested style words):

analog, AnalogRedmAF, F1.2 shallow depth of field, 35mm analog film photo, soft contrast, fine film grain, subtle halation, cinematic bokeh

python ~/.hermes/skills/create-image/scripts/create_image.py \
  "analog, AnalogRedmAF, F1.2 shallow depth of field, 35mm analog film photo, soft contrast, fine film grain, subtle halation, cinematic bokeh" \
  --image /path/to/photo.jpg --lora-scale 0.85 --steps 2

Locked film behavior

• The tool prepends the canonical RefControl/Analog Redmond film base and tuned grade automatically for photo film requests. Do not improvise or rewrite that base; provide only the requested extra style words.
• Film requests with no --image belong to create-image; its normal text-to-image route loads the analog LoRA.
• The Analog Redmond adapter is analog_redmond_fluxklein9b.safetensors, 9B-only. It is loaded on demand.
• Report that the Analog Redmond LoRA was loaded when it was applied.

Color grade vocabulary

Keep composition fixed with a --seed; change only the style words:

• warm: warm tones, golden / amber tone, warm Kodak Portra tones
• cool: cool tones, teal shadows, cyan-green cast, cool Fuji green-cyan tones, blue hour
• saturation: muted colors, desaturated, pastel palette, vivid colors
• contrast: soft contrast, low contrast, lifted blacks, faded blacks, washed-out, high contrast
• film stocks: Kodak Portra (warm skin), Fujicolor (green/cyan), Cinestill 800T (tungsten teal/red halation), Kodak Gold (warm nostalgic)
• monochrome: black and white, sepia, monochrome

Use --lora-scale to vary the whole analog effect; color words only change palette/grade.

Size and quality

Without an explicit aspect ratio, the tool follows source orientation and crops cover-style without distortion:

Source	Generation → final
Landscape	1248x832 → 1776x1184 (3:2)
Portrait	832x1248 → 1184x1776 (2:3)
Square	1024x1024 → 1440x1440

Use --aspect-ratio only when a user explicitly requests a different shape. Use --native-1080p only when they explicitly request native/high-quality output; it is much slower. Do not enable VAE tiling.

Operational constraints

• Runtime: /home/chihmin/models-work/flux2/.venv-rocm72; ROCm 7.2 / gfx1151. No model install/download is needed.
• Do not enable create-image-daemon.service; a warm FLUX daemon risks OOM.
• Do not stop/restart Qwen or Gemma MTP for photo editing.
• Do not use --fast-preview for RefControl; it has no matching 4B LoRAs.
• Avoid prompt words that push unwanted harsh or blown-out lighting unless explicitly requested: hard light, harsh light, overexposed.
• If native output OOMs, report it and fall back to the normal upscale path instead of stopping other services.