Turn a selfie into a woke profile with GPT Image 2.6

One selfie goes in. Out come 11 woke versions of you and a 4:5 collage titled "I MADE ME WOKE" for X, plus an MP4 where the card, the cat, the house lights and your protest interview move. It runs on GPT Image 2.6, the image model behind ChatGPT, through varg. The full set costs about 78 credits ($0.78): nine image edits at 7-8 credits, two GIFs at 7, and three shared tiles for free. One command:

python3 scripts/wokify.py --selfie me.jpg --name "Alex" --pronouns "he/they" --out wokify-out
python3 scripts/collage.py wokify-out --anim        # collage.jpg 2400x3000, collage.mp4, collage.gif

Example, made from one selfie on 2026-11-05:

• the collage (JPG, 4:5)
• the animated collage (MP4)
• the presenter-mode video (MP4, 42 s)

No API key? Paste the prompts from references/prompts.md into ChatGPT with your selfie attached, one tile per chat (section "ChatGPT route" below).

Rules before you start

• Use your own photo, or one you have the person's clear permission to use. Never a public figure, a child, or someone who did not agree. The output is a recognizable fake of a real face.
• Punch at habits and brands, not at people. The jokes are tote bags, oat lattes, land acknowledgments on LinkedIn and rainbow logos that vanish in July. Pride and trans flags appear as the things the characters wear, never as the punchline.
• No real brand logos. GPT Image 2.6 adds them by itself: a Patagonia patch on a fleece vest, New Balance "N" on sneakers, GT's on kombucha, the LinkedIn "in". The prompts forbid them; still check every tile (step 4).

Steps

1. Pick the selfie. It should be front-facing, sharp and evenly lit, about 1024 px or larger, with one face, nothing covering it, and a plain background. A smiling photo gives the best "smug" tiles. Ask for the person's name and pronouns: they go on the profile page and the call tile.
2. Price it first. python3 scripts/wokify.py --dry-run prints every tile and the total (~78 credits; --fresh remakes the shared cat, latte and house from their prompts: ~106).
3. Run it. wokify.py uploads the selfie (POST /v2/files) and runs the tiles four at a time on gpt_image_2_6. Each tile edits a hosted base image with the selfie as image 2, so the scene stays the same and only the face changes. The GIF steps then animate the tile's own still. About 2-3 minutes. Every job (id, credits, URL, prompt) goes to wokify-out/run.json.
4. Look at every face and every logo. Open each PNG at full size.
   • Face drifted? --fix-face <tile> runs a face-only pass on Flare with the selfie (8 credits) and keeps the old file as <tile>.before-face-fix.png.
   • A logo slipped in? --fix-logo <tile> --where "the sneakers" (7-8 credits).
   • Fixed a tile that has a GIF (social, card)? Re-animate it with --gif-only --only <tile>.
5. Build the collage. python3 scripts/collage.py wokify-out --anim. The rows are fixed (below); GIF tiles move in collage.mp4 and collage.gif; the JPG uses their first frame.
6. Optional: the presenter-mode video. python3 scripts/presenter.py wokify-out --music bed.mp3 writes a HyperFrames project. Render it with cd wokify-out/presenter && npx hyperframes@0.9.15 render -o final.mp4 --crf 20.
   • The video is 42 s at 1920x1080. The collage opens and scrolls for 9 s, then 11 slides of 3 s each show the current slide large, a NOW note, a NEXT preview, and every slide on a timeline with a rainbow highlight.
   • Music: any 42 s bed at 120 BPM keeps every cut on a beat. On varg: {"model": "eleven_music_v2_5", "prompt": "...", "duration": 42, "provider_options": {"elevenlabs": {"force_instrumental": true}}}. Pass the length as duration: a top-level music_length_ms is dropped silently and you get 30 s.
7. Post it. X shows a single 4:5 image uncropped in the feed. A GIF must be within 1280x1080, and phones only take 5 MB, which is why collage.gif is 720x900. The MP4 is the safer upload. A caption that fits: "I made me woke 🏳️‍🌈 One selfie, 11 prompts: my LinkedIn, my cat, my trading card, my protest era." Put links in the first reply: X shows posts with links less.

The 12 tiles

Collage rows (4:5, 2400x3000): BEFORE · MY FRIENDS / ME · MY CARD · MY LINKEDIN / MY FIGURE · MY PET · MY FAV DRINK · MY HOME / MY SOCIAL LIFE · MY FAVORITE GAME · MY COLLEAGUES.

The research behind two of the tiles was done on X on 2026-11-05:

• The #1 woke animal is the cat. JD Vance's "childless cat ladies" clip has 30.1M views and 29.6k likes, and a "WOKE CAT" meme has 110k likes. The runner-up, the pit bull, peaks around 2.5k.

• The woke friend group has five most-memed archetypes:

  • the they/them art student (teal hair, septum ring, pronoun pin);
  • the "performative male" feminist (mustache, cardigan, iced matcha, tote; 21k likes);
  • the wine-mom ally;
  • the masked DSA organizer;
  • the tie-dye boomer.

  The setting is a street protest, and the sign is the 11-04 "WOKE 3.0" meme (45-50k likes).

Sources are linked in references/prompts.md.

What we learned making it (measured on varg, 2026-11-05)

• Edit a base image, don't generate the scene. Sending [scene, selfie] with "replace the person with the person from image 2" keeps every label, layout and joke exact. Generating the whole scene from text with the selfie also works, but the text and layout come out different every time.
• Small faces come out mushy. In a 1088x608 image, faces in a group or storefront are about 25 px wide and lose your likeness. Fixes:
  • frame closer and raise image_size (for example 1536x1712 per panel);
  • for the sharpest faces, pin fal:openai/gpt-image-2.6/sunburst/edit at quality: "high";
  • Sunburst sharpens but drifts identity, so follow it with the Flare face-only pass (--fix-face).
• GIF edits of an existing image keep text readable. The card's stat grid, the TV chyron and the lawn sign stayed exact in every frame. Text the model has to invent inside a GIF still garbles past one word (see the gpt-image-gif skill).
• The model re-adds real logos from memory. It drew the LinkedIn "in" back onto a page that had a plain briefcase, until the prompt said "keep the briefcase logo, do not draw a letter logo".
• Frame sizes for GIFs. The frames share one sprite sheet capped at 3840 px on its long edge and 8.3 MP in total. 8 frames of 768x768, 10 of 1024x576 and 12 of 768x768 all pass.
• The files API refuses GIF uploads (they come back as .bin). Host GIFs elsewhere, or feed the still and let the model make the GIF.

ChatGPT route (no code)

1. Open ChatGPT, attach your selfie and, for a tile with a base, the base image. The links are in scripts/tiles.json → base.
2. Paste the tile's prompt from references/prompts.md. Replace {NAME} and {PRONOUNS}, and call the selfie "image 2".
3. For the GIF tiles, attach the result and paste the GIF prompt, asking for "a looping animated GIF" (@op7418 showed ChatGPT doing this on launch day). If your ChatGPT returns a still, use the varg route for the GIF tiles.
4. Save each result as <tile>.png or .gif in one folder and run collage.py on it. Without Python, any 4:5 grid tool works.

Files

• scripts/wokify.py: the runner. It uploads the selfie, runs the tiles and their GIFs in parallel, and handles the face and logo fixes, --dry-run, --only, --fresh and run.json. Standard library only.
• scripts/tiles.json: every tile with its base image URL, prompt, size and GIF settings, plus the fix prompts. Edit prompts here.
• scripts/collage.py: the 4:5 X collage as JPG, plus MP4 and GIF with --anim.
• scripts/presenter.py: the presenter-mode slideshow video as a HyperFrames project.
• references/prompts.md: every prompt verbatim, for ChatGPT, with the research links and the prompts that made the base images.

Related skills: gpt-image-2-6 (model guide, Flare vs Sunburst, prices), gpt-image-gif (GIF frame budget), varg-ai (API key, MCP).