photoreal-stills

Org pack for ultra-realistic stills: model routing, camera/lens prompting, multi-reference identity (incl. dogs/animals), upscale step, and a reject checklist so Layer B assets read as photographs — n

SKILL.md

GitHub

Photoreal Stills (org pack)

Use for brand, page, email, social, and paid still production when the claim is photoreal / photography-first (e.g. Blacksage Working-Dog Cinema).

Does not replace Plane B render skills — it gates how seats prompt, render, upscale, and accept stills.

Read with

Pack Role
skills/community/openmontage/.claude/skills/flux-best-practices/ FLUX.2 T2I/I2I, hex, multi-ref, no negatives
skills/community/visual-skills/image/ Prompt templates / model pick (prompt craft)
skills/integrations/ai-toolkit-local/ Mac primary: local FLUX.2-dev via Ostris ai-toolkit
skills/integrations/fal-media/ fal FLUX.2 max/pro commercial/API upgrade
skills/community/inference-sh/flux-image/ inference.sh FLUX path
skills/community/inference-sh/ai-image-generation/ Multi-model still render
skills/community/inference-sh/image-upscaling/ Final megapixel / print pass
skills/org/packs/production-artifacts/ Paths + production_status

Prerequisite: Design brief

Do not call image generation until a Design brief exists (see production-artifacts Design-before-production gate). The brief must include the exact generation prompt(s), brand hex/lighting locks, crop/aspect, and cited design packs (visual-skills/image, brand, email-design for headers, etc.). Production stills are rendered from the brief, not freestyled.

How “indistinguishable from reality” is achieved

Research + BFL guidance converge on a pipeline, not a magic adjective (“8k”, “ultra realistic”):

  1. Right final model — On Mac: FLUX.2-dev local (generation_used: local/flux-2-dev) via ai-toolkit when HF_TOKEN is set. Commercial API upgrade: FLUX.2 [max]/[pro] via fal. Use klein only for speed drafts. Cursor built-in image gen is draft / concept — not client-final. Local commercial Layer B requires license_basis: bfl-self-hosted-commercial or fal re-render.
  2. Photographic language — Prose subject + action + environment, then camera body, lens, aperture, film/sensor, lighting. BFL: camera/lens/film references outperform vague “professional photo” (FLUX.2 prompting guide).
  3. One job per image — Do not mix watercolor + cinematic + cyberpunk in one prompt; focus collapses realism (Artlist FLUX 2 practices).
  4. Positive-only — FLUX has no negatives; describe sharp focus, natural skin/fur, empty scene instead of “no blur / no CGI”.
  5. Multi-reference for identity — Lock dog/person/product with reference images; name each image’s role (identity / pose / lighting / environment). FLUX.2 supports multi-ref; use for kennel lines and recurring subjects (FLUX.2 product).
  6. Resolution ladder — Draft low → final high (up to ~4MP on FLUX.2 commercial) → optional upscale pack for print/retina.
  7. Human reject pass — Zoom 100%; fail plastic fur, wrong anatomy, CGI sheen, impossible DOF, text glitches (see checklist).

Prompt skeleton (FLUX finals)

[Subject with breed/age/markings] [pose/action]
[Environment — materials, time of day]
[Lighting — key direction, quality, color as hex when brand-locked]
Shot on [camera body], [focal length]mm at f/[aperture], [film stock or sensor look]
[Framing], shallow/deep DOF as needed
Photorealistic documentary photography, natural detail

Brand hex example (Blacksage): tan key #C4A35A on markings; void #070707.

Model routing (Plane B)

Stage Prefer Avoid as final
Ideation / mood FLUX.2 klein, Cursor GenerateImage Claiming production complete
Brand / page / email hero (Mac) FLUX.2-dev local via ai-toolkit-local + scripts/render-blacksage-stills.sh Unspecified “AI art” models
Brand / page / email hero (API) FLUX.2 max or pro via fal or inference.sh Draft-only gens
Paid stills Local FLUX.2-dev (licensed) or fal pro/max / GPT Image per ad-creative Cursor drafts
Upscale image-upscaling (Topaz / FLUX upscaler / Real-ESRGAN as pack allows) Stretching draft PNG in CSS

Record generation_used e.g. local/flux-2-dev or fal/flux-2-max + upscaler id + license_basis when local.

Dog / animal identity (kennel ventures)

Human character sheets are not enough. For recurring dogs:

  1. Collect 3–8 real or approved reference photos (angles: front, three-quarter, side; markings clear).
  2. Prompt with explicit markings (chest blaze, eyebrow dots, muzzle tan) + age/sex/build.
  3. Multi-ref: image 1 = identity lock; image 2 = pose; image 3 = lighting/environment.
  4. Reject if markings drift, eyes look glass/CGI, coat looks wet plastic, or proportions cartoonize.
  5. Same seed + refs when iterating a locked hero.

Photoreal reject checklist (required before production_status: complete)

Zoom to 100% on the delivery file:

  • Reads as a photograph, not illustration / 3D render / AI poster
  • No plastic/waxy skin or fur; natural micro-texture visible
  • Eyes: catchlights plausible; no dead “glass” or doubled pupils
  • Anatomy OK (limbs, ears, teeth if visible) — no melted digits/claws
  • Lighting physics consistent (one key story; shadows match)
  • Depth of field matches stated aperture (background blur not fake smear)
  • No random text, watermarks, or logo glitches
  • Brand hex / grade not purple glow / cream terracotta AI defaults
  • For recurring subject: markings match refs
  • Resolution suitable for lease (email header ≥1200px wide; heroes ≥1920 on long edge or upscaled)

Fail any → regenerate or I2I fix; do not ship.

Seat duties

Seat Duty
brand-designer Owns photoreal stills; runs checklist; upscales finals
web-designer UI stills via same pack when leased
paid-media-manager Ad stills; same checklist
creative-director Prompt + photoreal QA before merge
verifier Spot-check checklist + files on disk for claimed complete stills

Handoff fields (add when stills produced)

generation_used: fal/flux-2-max   # or inference-sh/...
photoreal_qa: pass | fail
photoreal_notes: ""               # what was checked / why fail
reference_paths: []               # multi-ref inputs if any

Anti-patterns

  • Shipping Cursor draft gens as Layer B finals
  • “8k ultra detailed masterpiece” keyword salad
  • Negative-prompt habits on FLUX
  • Completing production without zoom QA
  • Treating Figma mock as the still (export required — see production-artifacts)