Free-tier image generation · recon

Same prompt, six tools, 2026.

Pick the lineup for the recreation pilot. One identical prompt sent to every free image generator we could reach without a paywall today, side-by-side with the 2022-2023 MidJourney original Thomas already made.

test prompt · ⟶ · 6 generators · ⟶ · 1 sanity check

MidJourney '22 archive

Thomas's original — MJ v4, late 2022. The reference Thomas is trying to reach again.

1664×1664 · 3.7 MB archive

Google Gemini web

Logged in as Curio. Cinematic full-scene with cherry blossom + pagoda. ~25s/image, no visible counter.

1024×559 · 1.3 MB working

AI Studio nano banana

Google's gemini-2.5-flash-image. Free tier. ~30s/image, ~1300 tokens. Character-portrait aesthetic.

1024×1024 · 2.3 MB working

HF FLUX merged

multimodalart/FLUX.1-merged on a *.hf.space direct host. ZeroGPU queue is flaky but anonymous-tolerable.

1024×1024 · 1.0 MB flaky

mflux local disk-blocked

Apple Silicon native, totally local, no quota. Install OK; every non-gated model in mflux exceeds the available host disk.

[ — ]

Followed by: tried z-image-turbo (Tongyi-MAI, non-gated) — partially downloaded 21 GB before filling the disk. Tried fibo-lite (briaai, gated 401). Tried qwen (~40 GB, non-gated but too big). Tried flux2-klein-base-4b (BFL, non-gated) — partial 8 GB before disk-full again.

Needs ~25 GB free disk (host is 460 GB / 93% used); revisit after Thomas frees host capacity, or run mflux on a different machine.

install OK · disk full action needed

ChatGPT gpt-image-1

Plus account, browser-driven via Chrome MCP. ~70s end-to-end. Cherry-blossom scene with painterly samurai.

1254×1254 · 2.6 MB working

Decision point

Approve this lineup?

If this set looks right, the pilot recreates 35 of Thomas's prompts using Gemini + AI Studio + HF FLUX. Swap a column if one looks weaker than expected, or request a second prompt as a tiebreaker before committing.

// these are visual placeholders — say the word in chat and I'll act on the decision