astro-knots
Mode

← All notes

explorations

AI Photo-to-Illustration Transform for Arthouse

Exploring how to take a real client photograph — sometimes NSFW-sensitive — and return an obviously AI-generated illustrative version (anime, painterly, or stylized) that is safe to publish on the public arthouse-site portfolio.

Draft

Parent spec: Polish Pass for Arthouse Site · Maintain an Image Heavy Portfolio Site Status: Draft — destination unclear, need to evaluate tools

1. The problem

The arthouse client's body of work includes photography that is racy / NSFW. To showcase that work publicly without exposing the original imagery, we want a pipeline that:

  1. Takes a source photograph as input

  2. Returns an output that is clearly AI-generated (so no reasonable viewer mistakes it for the original photo)

  3. Preserves the composition, mood, and subject pose of the source — these are what the artist wants to be hired for

  4. Has a consistent visual style across a shoot (multiple transformed images read as a series, not random outputs)

  5. Is fast/cheap enough to iterate at the volume of a portfolio shoot (50–200 images)

The output style should read as illustration, not photo-realism — anime, painterly, ink-and-wash, or graphic novel. Photo-realistic style transfer would defeat the privacy purpose.

2. Why this is its own document

The polish pass on the home page can ship without solving this. But this is a load-bearing capability for the portfolio surface — without it, large swaths of the client's actual work cannot be shown.

3. Candidate tools

3.1 Ideogram v3 (already wired)

  • Already used by generate consistent og images skill

  • IDEOGRAM_API_KEY already in ~/.secrets

  • Supports style reference images and character reference (consistency across a series)

  • Open question: Does Ideogram do meaningful image-to-image conditioning on a source photo, or only style-reference (where the source guides aesthetic, not composition)? Need to verify against current API docs.

3.2 Replicate

  • Many models, pay-per-call, no infra

  • Strong candidates:

    • flux-dev / flux-schnell with image conditioning + LoRA for anime style

    • sdxl + ControlNet (pose / canny / depth) — best for composition fidelity

    • Dedicated anime models: animagine-xl, anything-v5, counterfeit-xl

  • Trade-off: more control, more knobs, longer iteration loop

3.3 Stability AI direct API

  • SD 3 / SDXL with strength-controlled img2img

  • Style fidelity good, composition-preservation requires ControlNet

  • More setup than Replicate

3.4 Local (ComfyUI / Automatic1111)

  • Full control, zero per-call cost after setup

  • High overhead for one-off creative pipelines

  • Probably overkill for a single-client portfolio

4. The composition-vs-style tradeoff

Two failure modes to design against:

  • Too faithful to source: output reads as "a photo with a filter" — the NSFW essence isn't actually obscured, just stylized. Defeats the privacy purpose.

  • Too liberal with source: output reads as "generic anime girl" — loses the specific pose, composition, lighting that makes the client's work theirs.

The right control point is probably ControlNet pose/depth on a strongly anime-styled base model — keeps the silhouette and composition, replaces the rendering.

5. Series consistency

A portfolio shoot is 20–50 images of the same subject. They need to read as a series. Options:

  • Train a small LoRA on the subject (overkill for one-off)

  • Use a character reference image across all calls (Ideogram supports this)

  • Use the same seed + prompt scaffold across the batch with only pose changing

  • Manually curate the most consistent outputs and discard outliers (most pragmatic)

6. Privacy-of-the-pipeline question

If we send the source photos to a third-party API (Ideogram, Replicate, Stability), the source photo leaves the local machine. Depending on the provider's data-retention policy, this may or may not be acceptable for genuinely sensitive material. Worth checking each provider's policy before this is more than a prototype.

Local-only inference (option 3.4) is the only way to fully eliminate this risk.

7. Next moves

  1. Pick one candidate to prototype against — recommend Ideogram first because it's already wired and the lift to test is minimal. If style-reference-only is the constraint, fall back to Replicate flux+ControlNet.

  2. Test on a single non-sensitive source photo — verify composition preservation and "obviously AI" output before exposing any real NSFW source.

  3. Document the data-retention finding for whichever provider we pick.

  4. Sketch the directory flow — private/transform-source/<name>.jpg → script call → public/images/ai-rendered/<name>.webp.

  5. Decide where the transform script lives — sites/arthouse-site/scripts/? Or a shared utility?

This is the exploration phase. No commitments yet.

8. References

  • Polish Pass for Arthouse Site — parent

  • generate consistent og images skill — current Ideogram usage

  • crawl fetch ingest skill — ~/.secrets API key pattern