Skip to content
Python

README: why Create Image runs on FLUX 3 Klein, with the comparison sheet; runs 13-14 (34 runs, 646 of 669)

neuregex299c32e8012b
diff --git a/README.md b/README.md
index 4c0fddf..2d54595 100644
--- a/README.md
+++ b/README.md
@@ -19,22 +19,24 @@ browser ──▶ web/ + server/ (Node, Comfy SDK) ──▶ gateway :8000 ─
| | |
|---|---|
-| Apps | 112, all *ready*: 33 apps with a workflow of their own, the 5 IC-LoRA apps (`genjutsu`, `reframe`, `sdr-hdr`, `mixed-media`, `dub`), `morph`, 71 template apps (a fixed prompt over one of those pipelines, written by `tools/template_apps.py`) and 2 model-free smoke apps. The 110 that use models each have a lab example made on a GPU |
+| Apps | 114: 113 *ready* (34 apps with a workflow of their own, the 5 IC-LoRA apps (`genjutsu`, `reframe`, `sdr-hdr`, `mixed-media`, `dub`), `morph`, 71 template apps (a fixed prompt over one of those pipelines, written by `tools/template_apps.py`) and 2 model-free smoke apps) and `moss-image` as a draft. The 112 that use models each have a lab example made on a GPU |
| Open weights | No API or partner nodes. Every model is a file you download, listed in its app's `app.json` with its Hugging Face repo, size and license; the cloud text-encoder branch of Lightricks' examples is removed |
| Workflows | Built from official sources by `tools/build_apps.py` (ComfyUI's templates, Lightricks' ComfyUI-LTXVideo examples, Comfy's `comfy-reshoot`); the template apps reuse those bases and the two smoke apps are hand-written. `tools/import_workflow.py` drafts a new app from any other ComfyUI workflow. All accepted by ComfyUI's validator: ComfyUI `e3bee7f` (0.42.0), ComfyUI-LTXVideo `44c8a36`, ComfyUI-CrossViewWarp `fbb1e93` |
-| GPU runs | 32 lab runs on Modal from 20 to 24 Oct (L40S and H100), 607 of 630 cases ok. Every example in the web UI comes from one of them, with its time and peak VRAM. Run by run, and what each one changed: [`docs/lab-runs.md`](docs/lab-runs.md) |
+| GPU runs | 34 lab runs on Modal from 20 to 24 Oct (L40S and H100), 646 of 669 cases ok. Every example in the web UI comes from one of them, with its time and peak VRAM. Run by run, and what each one changed: [`docs/lab-runs.md`](docs/lab-runs.md) |
| End to end | Run from the browser on 23 Oct, web UI → server → gateway on Modal → ComfyUI on an H100: `morph` 160.8 s from the button to the video (128.7 s of it on the GPU), `mixed-media` 149 s, `dub` 72 s. On 24 Oct through the hosted demo (`infra/web_app.py`, the same web app on Modal behind a passcode): `face-swap` 25.2 s on the GPU. The model-free smoke apps cover the same path in CI |
-| LoRA lab | 268 cases: one showcase per app (110), one per effect (24), one per look of `mixed-media` (16) and one per transformation of `morph` (15), 36 references the others start from, 6 candidate portraits for those references (`refx-*`), 12 style-drift experiments (`mmx-*`), 36 template experiments (`tplx-*`), 2 runs of `reshoot` without its Turbo LoRA and 2 soundtrack candidates for the demo video; each experiment has its verdict in the note |
+| LoRA lab | 300 cases: one showcase per app (111), one per effect (24), one per look of `mixed-media` (16) and one per transformation of `morph` (15), 36 references the others start from, 6 candidate portraits for those references (`refx-*`), 12 style-drift experiments (`mmx-*`), 36 template experiments (`tplx-*`), 2 runs of `reshoot` without its Turbo LoRA, 30 image-model comparisons (`cmp-*`, below) and 4 soundtrack candidates for the demo video; each experiment has its verdict in the note |
| CI | `.github/workflows/validate.yml` on every push: the manifest validator, the tests, the type check, the generated packs and VRAM map, and a CPU-only ComfyUI that accepts every workflow and the prompt of every lab case (about 3 minutes) |
## Apps
| App | Rebuilds | Model | On a 24 GB card | Status |
|---|---|---|---|---|
-| `z-image` | Create Image | Z-Image Turbo (int8) | fits whole, 12.2 GB | ready |
+| `create-image` | Create Image | FLUX 3 Image Klein (int8) + BFL's 9B prompt upsampler (an on/off toggle); a layout table in JSON can follow the prompt | fits whole, 17.8 GB | ready |
+| `z-image` | Create Image, fast drafts | Z-Image Turbo (int8); the lab makes its reference photos with it | fits whole, 12.2 GB | ready |
+| `moss-image` | Create Image with a layout | MOSS-Image SFT (int8), Apache-2.0; also takes a layout JSON with a box per text and object | fits whole, 13.0 GB | draft |
| `styles` | Image styles | Krea 2 Turbo (int8) + Krea's nine official style LoRAs, picked from a list | fits whole, 19.4 GB | ready |
| `style-reference` | Style transfer | Krea 2 Turbo (int8) + style-reference LoRA | fits whole, 19.4 GB | ready |
-| `cinema-studio` | Cinematic Cameras | Z-Image Turbo + camera/lens vocabulary | fits whole, 12.2 GB | ready |
+| `cinema-studio` | Cinematic Cameras | FLUX 3 Image Klein + camera/lens vocabulary, 16:9 | fits whole, 11.0 GB | ready |
| `edit` | Inpaint / Edit | FLUX 3 Image Klein (int8) | fits whole, 11.0 GB | ready |
| `relight` | Relight | Qwen Image Edit 2509 (fp8) + Relight and Lightning LoRAs, light presets | tight: 21.5 GB largest phase | ready |
| `upscale`, `upscale-video` | Upscale | SeedVR2 3B (int8) | fits whole, 4.0 GB | ready |
@@ -67,13 +69,28 @@ browser ──▶ web/ + server/ (Node, Comfy SDK) ──▶ gateway :8000 ─
| `talking` | Lipsync / Speak | LTX-2.5 audio-to-video (int8) | tight: 21.5 GB largest phase | ready |
| `smoke`, `smoke-invert` | — | none | — | ready |
+### Which image model
+
+Create Image ran on Z-Image Turbo until 24 Oct. Artificial Analysis' text-to-image arena (October 2026) ranks FLUX 3
+Image Klein first among open-weights models (1043 Elo; Qwen-Image-2.1 1042, Bria's Fibo 2.0 1039, Krea 2 up to 1034)
+and Z-Image Turbo at 936, and Klein is the file the edit apps already download. The lab ran six prompts on five
+models with the same seed (run `20261024-230058-l40s`, below): Klein kept the counts (three apples, five lemons) where
+Z-Image drew four apples, wrote the label Krea 2 left blank, kept the top-down view Krea missed, and read a JSON prompt
+as a scene where Z-Image printed its keys on the sign. So `create-image`, `cinema-studio`, `ai-influencer` and `meme`
+run on Klein. BFL's 9B upsampler is a toggle in Create Image (on by default: it turns a short prompt into a full
+caption, about 8 s more). Z-Image stays for fast drafts and for the lab's reference photos, so earlier examples keep
+their inputs. MOSS-Image (Apache-2.0) is in as a draft for layout prompts. Fibo 2.0, trained on JSON prompts, needs a
+newer ComfyUI than the one pinned here and its weights are licensed for non-commercial use only.
+
+![Six prompts (a portrait, a product with text, counting, a JSON prompt, a poster, a watercolor) on Z-Image Turbo, FLUX 3 Klein with and without the upsampler, Krea 2 Turbo and MOSS-Image, with GPU seconds and peak VRAM under each](docs/media/image-models.jpg)
+
### Template apps
Higgsfield's catalog is mostly fixed prompts over a few pipelines, and so is this: each of these apps is one of
Freefield's pipelines with its prompt baked in, selects that fill parts of it, and an "Extra direction" box appended
to it. The bases: `edit` (FLUX 3 Image Klein, one reference image), `edit` with a second reference (a garment, another
person; for `face-swap`, MediaPipe's face oval from core ComfyUI nodes masks the edit and the face photo is Klein's only
-reference; for `expand-image`, the photo gets a border that only Klein draws), `z-image` (text to image), `video`
+reference; for `expand-image`, the photo gets a border that only Klein draws), `create-image` without its upsampler (Klein text to image), `video`
(LTX-2.5 image-to-video with its own audio), `morph` (Klein draws the last frame from a fixed instruction, LTX-2.5
and the transition LoRA move the photo into it), `effects`' Wan 2.1 pipeline with one effect LoRA fixed (joyfox's
Bullet Time; for `bullet-time-white` and `-splash`, Klein redraws the photo first) and `relight`'s
@@ -109,7 +126,7 @@ command that makes every example. Add a line to the table there, not a folder by
| `paint-app` | Paint App | `edit` | your photo |
| `mural` | Signboard | `edit` | your photo |
| `micro-beasts` | Surrounded by Animals | `edit` | your photo, animals |
-| `meme` | Meme Generator | `z-image` | scene, top text, bottom text |
+| `meme` | Meme Generator | `create-image` | scene, top text, bottom text |
| `j-magazine` | J-Magazine | `edit` | your photo |
| `cafe-60s` | 60s Cafe | `edit` | your photo |
| `victory-card` | Victory Card | `edit` | your photo |
@@ -147,7 +164,7 @@ command that makes every example. Add a line to the table there, not a folder by
| `rap-video` | Rap God | `video` | your photo |
| `idol` | Idol | `video` | your photo |
| `expand-image` | Expand Image | `edit` | your photo, expand |
-| `ai-influencer` | AI Influencer | `z-image` | who, niche |
+| `ai-influencer` | AI Influencer | `create-image` | who, niche |
| `bullet-time-scene` | Bullet Time Scene | `effects` | your photo, who or what is in the photo |
| `bullet-time-white` | Bullet Time White | `edit + effects` | your photo, who or what is in the photo |
| `bullet-time-splash` | Bullet Time Splash | `edit + effects` | your photo, who or what is in the photo |
@@ -183,7 +200,7 @@ modal deploy infra/modal_app.py # FREEFIELD_GPU=A10G to try th
npm install
# in .env (made by step 1): COMFY_BASE_URL = the URL printed by step 3
npm run serve # http://127.0.0.1:3000
-npm run app -- z-image --set prompt="a red bicycle against a white wall" # or from the terminal
+npm run app -- create-image --set prompt="a red bicycle against a white wall" # or from the terminal
```
On Windows PowerShell, environment variables go before the command as `$env:FREEFIELD_GPU = "A10G"; modal deploy ...`.
@@ -245,7 +262,7 @@ the CLI until you set `FREEFIELD_DRAFTS=1` (or pass `--draft` to `npm run app`).
runs on them with a fixed seed. Each app has a *showcase* case, so one command makes an example of every app:
```sh
-modal run --detach lab/lab_modal.py --showcase # one example per app + the references they need (138 cases)
+modal run --detach lab/lab_modal.py --showcase # one example per app + the references they need (139 cases)
modal run lab/lab_modal.py --pull <run_id> # only if the terminal closed before the end: copy back and publish
```
diff --git a/docs/lab-runs.md b/docs/lab-runs.md
index ed3c8ea..c25ecf4 100644
--- a/docs/lab-runs.md
+++ b/docs/lab-runs.md
@@ -3,7 +3,7 @@
Every GPU run of the lab (`lab/lab_modal.py`) from 20 to 24 Oct 2026, in order. Each run is a folder
`lab/results/<run id>` with `results.jsonl` (one row per case: status, seconds, peak VRAM), `run.json` (GPU, start
time) and `review.html`. Peaks are the highest of the run; on an 80 GB H100 ComfyUI can keep the models of earlier
-steps loaded, so a peak there is not the least an app needs. 32 runs, 607 of 630 cases ok: 22 of the 23 failures are
+steps loaded, so a peak there is not the least an app needs. 34 runs, 646 of 669 cases ok: 22 of the 23 failures are
the first IC-LoRA run, before the gated licenses were accepted, and the other is `recast` in the first run.
| Run | GPU | What ran | Cases ok | Longest case | Peak |
@@ -40,6 +40,8 @@ the first IC-LoRA run, before the gated licenses were accepted, and the other is
| `20261024-181216-h100` | H100 | Queue job 10: the runner in a teal and black kit (the red and white one had a brand logo) | 3 / 3 | 189 s | 39.9 GB |
| `20261024-182254-h100` | H100 | Queue job 11: Klein removes the logo from the runner's top; Bullet Time Splash and Hero run re-made | 4 / 4 | 192 s | 40.3 GB |
| `20261024-192222-h100` | 3 × H100 | Queue job 12: brand logos out of the road photo, the athlete, Mascot and Style Snap; `reshoot` without the Turbo LoRA | 16 / 16 | 346 s | 59.0 GB |
+| `20261024-230058-l40s` | 3 × L40S | Queue job 13: six prompts on five image models (`cmp-*`); Cinema Studio, AI Influencer and Meme on FLUX 3 Klein; `whats-next` from the lake | 37 / 37 | 37 s | 28.7 GB |
+| `20261024-232228-l40s` | L40S | Queue job 14: two soundtracks of two minutes for the demo video | 2 / 2 | 161 s | 37.0 GB |
A run with several GPUs (`--shards N`) splits the cases into N groups that run at the same time and comes back as
one folder. The cases a run lists include the reference images, clips and songs that the others read.
@@ -90,6 +92,14 @@ one folder. The cases a run lists include the reference images, clips and songs
plain steps were not better than the 8 Turbo steps, softer with the same holes, and took 249 s and 346 s against
190 s. The app keeps the Turbo LoRA and shows the license note next to it.
+- **The image model.** Run 13 put six prompts on Z-Image Turbo, FLUX 3 Klein (with and without BFL's 9B prompt
+ upsampler), Krea 2 Turbo and MOSS-Image, same seed. Klein followed the prompts best: three apples and five lemons
+ where Z-Image drew four apples, the label text Krea 2 left out, the top-down view Krea missed, and a JSON prompt read
+ as a scene where Z-Image printed its keys on the sign. Create Image, Cinema Studio, AI Influencer and Meme moved to
+ Klein; the upsampler added about 8 s on an L40S. The comparison sheet is `docs/media/image-models.jpg`.
+- **Soundtrack.** With empty lyrics MiniMax Music 3 made 20 s pieces (job 13); listing the song's sections as tags
+ gave 130 s, and Stable Audio 3's music category gave 120 s with a quiet intro and a clean ending (job 14).
+
## End to end from the browser (23 Oct)
Web UI → server → gateway on Modal → comfy-api-proxy → ComfyUI on an H100: `morph` (Astronaut) 160.8 s from the button
diff --git a/docs/media/image-models.jpg b/docs/media/image-models.jpg
new file mode 100644
index 0000000..a993bb0
Binary files /dev/null and b/docs/media/image-models.jpg differ
diff --git a/lab/cases.json b/lab/cases.json
index 9d63737..5609164 100644
--- a/lab/cases.json
+++ b/lab/cases.json
@@ -54,75 +54,75 @@
"note": "Transition LoRA from 2.3 on 2.5. The morph should be smooth and stay on her face."},
{"id": "demo-music-c", "app": "music", "values": {"style": "Global metadata: instrumental electronic pop, 120 bpm, 4/4, bright and confident, for a product demo video. Vocal details: no vocals, no voice, no choir. Arrangement: punchy drums, driving synth bass, bright plucked arpeggios, warm pads; builds from a sparse intro to a full chorus, a short breakdown, a last big chorus and a clean ending.", "lyrics": "[intro]\n\n[verse]\n\n[chorus]\n\n[verse]\n\n[chorus]\n\n[bridge]\n\n[chorus]\n\n[outro]", "seconds": 130},
- "note": "Soundtrack candidate for the demo video: empty lyrics gave 20 s pieces (demo-music-a/b), so this one lists the song's sections as tags, the only structure MiniMax Music 3 follows."},
+ "note": "Soundtrack candidate for the demo video: empty lyrics gave 20 s pieces (demo-music-a/b), so this one lists the song's sections as tags, the only structure MiniMax Music 3 follows. Verdict (run 20261024-232228-l40s): 130 s, steady, a short breakdown near 0:55."},
{"id": "demo-music-d", "app": "sfx", "values": {"description": "instrumental electronic pop music bed, 120 bpm, punchy drums, driving synth bass, bright plucked arpeggios, warm pads, confident and modern, builds and ends cleanly", "category": "Music", "seconds": 120},
- "note": "Soundtrack candidate for the demo video with Stable Audio 3 (the sfx app's Music category)."},
+ "note": "Soundtrack candidate for the demo video with Stable Audio 3 (the sfx app's Music category). Verdict (run 20261024-232228-l40s): 120 s with a quiet intro and a clean ending; the draft video uses it."},
{"id": "ref-song", "showcase": true, "app": "music", "values": {"style": "warm acoustic folk song, clear female lead vocal, gentle guitar, 92 bpm", "lyrics": "[verse]\nMorning light across the window\nCoffee warm and the city slow\n[chorus]\nSing it out, sing it out\nEvery little thing we know", "seconds": 10},
"note": "Vocal reference for the talking avatar (MiniMax Music 3, original lyrics)."},
{"id": "cmp-fisher-zi", "app": "z-image", "values": {"prompt": "Candid photo of a 60-year-old fisherman from Veracruz mending a fishing net on a wooden dock at golden hour, weathered skin, gray stubble, faded blue work shirt, gulls in the background, shot on 35mm film", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five read as photos; Klein has the most skin and net detail, Z-Image is flatter."},
{"id": "cmp-fisher-kl", "app": "create-image", "values": {"prompt": "Candid photo of a 60-year-old fisherman from Veracruz mending a fishing net on a wooden dock at golden hour, weathered skin, gray stubble, faded blue work shirt, gulls in the background, shot on 35mm film", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five read as photos; Klein has the most skin and net detail, Z-Image is flatter."},
{"id": "cmp-fisher-klx", "showcase": true, "app": "create-image", "values": {"prompt": "Candid photo of a 60-year-old fisherman from Veracruz mending a fishing net on a wooden dock at golden hour, weathered skin, gray stubble, faded blue work shirt, gulls in the background, shot on 35mm film", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five read as photos; Klein has the most skin and net detail, Z-Image is flatter."},
{"id": "cmp-fisher-kr", "app": "styles", "values": {"prompt": "Candid photo of a 60-year-old fisherman from Veracruz mending a fishing net on a wooden dock at golden hour, weathered skin, gray stubble, faded blue work shirt, gulls in the background, shot on 35mm film", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five read as photos; Klein has the most skin and net detail, Z-Image is flatter."},
{"id": "cmp-fisher-ms", "app": "moss-image", "values": {"prompt": "Candid photo of a 60-year-old fisherman from Veracruz mending a fishing net on a wooden dock at golden hour, weathered skin, gray stubble, faded blue work shirt, gulls in the background, shot on 35mm film", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five read as photos; Klein has the most skin and net detail, Z-Image is flatter."},
{"id": "cmp-perfume-zi", "app": "z-image", "values": {"prompt": "Studio product photo of a square glass perfume bottle with a gold cap and a black square label that reads \"NOIR\" in white capitals, standing on wet black slate, rim light from behind, water droplets on the glass", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Krea 2 left the label blank; the other four wrote NOIR."},
{"id": "cmp-perfume-kl", "app": "create-image", "values": {"prompt": "Studio product photo of a square glass perfume bottle with a gold cap and a black square label that reads \"NOIR\" in white capitals, standing on wet black slate, rim light from behind, water droplets on the glass", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Krea 2 left the label blank; the other four wrote NOIR."},
{"id": "cmp-perfume-klx", "app": "create-image", "values": {"prompt": "Studio product photo of a square glass perfume bottle with a gold cap and a black square label that reads \"NOIR\" in white capitals, standing on wet black slate, rim light from behind, water droplets on the glass", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Krea 2 left the label blank; the other four wrote NOIR."},
{"id": "cmp-perfume-kr", "app": "styles", "values": {"prompt": "Studio product photo of a square glass perfume bottle with a gold cap and a black square label that reads \"NOIR\" in white capitals, standing on wet black slate, rim light from behind, water droplets on the glass", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Krea 2 left the label blank; the other four wrote NOIR."},
{"id": "cmp-perfume-ms", "app": "moss-image", "values": {"prompt": "Studio product photo of a square glass perfume bottle with a gold cap and a black square label that reads \"NOIR\" in white capitals, standing on wet black slate, rim light from behind, water droplets on the glass", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Krea 2 left the label blank; the other four wrote NOIR."},
{"id": "cmp-table-zi", "app": "z-image", "values": {"prompt": "Top-down photo of a wooden kitchen table: exactly three red apples on the left, a blue ceramic bowl holding five lemons in the center, a folded newspaper on the right, a silver fork lying across the bowl, soft morning light", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image drew four apples and Krea 2 an angled view; Klein (both) and MOSS kept three apples, five lemons and the top-down view."},
{"id": "cmp-table-kl", "app": "create-image", "values": {"prompt": "Top-down photo of a wooden kitchen table: exactly three red apples on the left, a blue ceramic bowl holding five lemons in the center, a folded newspaper on the right, a silver fork lying across the bowl, soft morning light", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image drew four apples and Krea 2 an angled view; Klein (both) and MOSS kept three apples, five lemons and the top-down view."},
{"id": "cmp-table-klx", "app": "create-image", "values": {"prompt": "Top-down photo of a wooden kitchen table: exactly three red apples on the left, a blue ceramic bowl holding five lemons in the center, a folded newspaper on the right, a silver fork lying across the bowl, soft morning light", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image drew four apples and Krea 2 an angled view; Klein (both) and MOSS kept three apples, five lemons and the top-down view."},
{"id": "cmp-table-kr", "app": "styles", "values": {"prompt": "Top-down photo of a wooden kitchen table: exactly three red apples on the left, a blue ceramic bowl holding five lemons in the center, a folded newspaper on the right, a silver fork lying across the bowl, soft morning light", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image drew four apples and Krea 2 an angled view; Klein (both) and MOSS kept three apples, five lemons and the top-down view."},
{"id": "cmp-table-ms", "app": "moss-image", "values": {"prompt": "Top-down photo of a wooden kitchen table: exactly three red apples on the left, a blue ceramic bowl holding five lemons in the center, a folded newspaper on the right, a silver fork lying across the bowl, soft morning light", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image drew four apples and Krea 2 an angled view; Klein (both) and MOSS kept three apples, five lemons and the top-down view."},
{"id": "cmp-json-zi", "app": "z-image", "values": {"prompt": "{\"scene\": \"a small taqueria at night in Mexico City\", \"subjects\": [{\"who\": \"a taquero in a white apron\", \"action\": \"slicing al pastor from the trompo\", \"position\": \"center\"}, {\"who\": \"two customers\", \"action\": \"waiting at the counter\", \"position\": \"left\"}], \"text\": [{\"content\": \"TACOS EL GÜERO\", \"where\": \"neon sign above the counter\", \"color\": \"red\"}], \"camera\": {\"lens\": \"35mm\", \"angle\": \"eye level\", \"depth_of_field\": \"shallow\"}, \"lighting\": \"warm tungsten light inside, cool blue street light outside\", \"style\": \"documentary photo, natural colors\"}", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image printed JSON keys (\"who\", \"where\", \"above the conter\") on the sign; the other four read it as a scene with TACOS EL GÜERO in red neon."},
{"id": "cmp-json-kl", "app": "create-image", "values": {"prompt": "{\"scene\": \"a small taqueria at night in Mexico City\", \"subjects\": [{\"who\": \"a taquero in a white apron\", \"action\": \"slicing al pastor from the trompo\", \"position\": \"center\"}, {\"who\": \"two customers\", \"action\": \"waiting at the counter\", \"position\": \"left\"}], \"text\": [{\"content\": \"TACOS EL GÜERO\", \"where\": \"neon sign above the counter\", \"color\": \"red\"}], \"camera\": {\"lens\": \"35mm\", \"angle\": \"eye level\", \"depth_of_field\": \"shallow\"}, \"lighting\": \"warm tungsten light inside, cool blue street light outside\", \"style\": \"documentary photo, natural colors\"}", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image printed JSON keys (\"who\", \"where\", \"above the conter\") on the sign; the other four read it as a scene with TACOS EL GÜERO in red neon."},
{"id": "cmp-json-klx", "app": "create-image", "values": {"prompt": "{\"scene\": \"a small taqueria at night in Mexico City\", \"subjects\": [{\"who\": \"a taquero in a white apron\", \"action\": \"slicing al pastor from the trompo\", \"position\": \"center\"}, {\"who\": \"two customers\", \"action\": \"waiting at the counter\", \"position\": \"left\"}], \"text\": [{\"content\": \"TACOS EL GÜERO\", \"where\": \"neon sign above the counter\", \"color\": \"red\"}], \"camera\": {\"lens\": \"35mm\", \"angle\": \"eye level\", \"depth_of_field\": \"shallow\"}, \"lighting\": \"warm tungsten light inside, cool blue street light outside\", \"style\": \"documentary photo, natural colors\"}", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image printed JSON keys (\"who\", \"where\", \"above the conter\") on the sign; the other four read it as a scene with TACOS EL GÜERO in red neon."},
{"id": "cmp-json-kr", "app": "styles", "values": {"prompt": "{\"scene\": \"a small taqueria at night in Mexico City\", \"subjects\": [{\"who\": \"a taquero in a white apron\", \"action\": \"slicing al pastor from the trompo\", \"position\": \"center\"}, {\"who\": \"two customers\", \"action\": \"waiting at the counter\", \"position\": \"left\"}], \"text\": [{\"content\": \"TACOS EL GÜERO\", \"where\": \"neon sign above the counter\", \"color\": \"red\"}], \"camera\": {\"lens\": \"35mm\", \"angle\": \"eye level\", \"depth_of_field\": \"shallow\"}, \"lighting\": \"warm tungsten light inside, cool blue street light outside\", \"style\": \"documentary photo, natural colors\"}", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image printed JSON keys (\"who\", \"where\", \"above the conter\") on the sign; the other four read it as a scene with TACOS EL GÜERO in red neon."},
{"id": "cmp-json-ms", "app": "moss-image", "values": {"prompt": "{\"scene\": \"a small taqueria at night in Mexico City\", \"subjects\": [{\"who\": \"a taquero in a white apron\", \"action\": \"slicing al pastor from the trompo\", \"position\": \"center\"}, {\"who\": \"two customers\", \"action\": \"waiting at the counter\", \"position\": \"left\"}], \"text\": [{\"content\": \"TACOS EL GÜERO\", \"where\": \"neon sign above the counter\", \"color\": \"red\"}], \"camera\": {\"lens\": \"35mm\", \"angle\": \"eye level\", \"depth_of_field\": \"shallow\"}, \"lighting\": \"warm tungsten light inside, cool blue street light outside\", \"style\": \"documentary photo, natural colors\"}", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): Z-Image printed JSON keys (\"who\", \"where\", \"above the conter\") on the sign; the other four read it as a scene with TACOS EL GÜERO in red neon."},
{"id": "cmp-poster-zi", "app": "z-image", "values": {"prompt": "A minimalist concert poster: the title \"MIDNIGHT TACOS\" in huge bold letters at the top, \"LIVE · FRIDAY 9 PM\" in smaller letters below it, an illustration of a neon taco truck at night, risograph print texture, only hot pink and teal", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five wrote both lines; Krea 2 and MOSS lost the two-color risograph look; Klein with the upsampler is closest."},
{"id": "cmp-poster-kl", "app": "create-image", "values": {"prompt": "A minimalist concert poster: the title \"MIDNIGHT TACOS\" in huge bold letters at the top, \"LIVE · FRIDAY 9 PM\" in smaller letters below it, an illustration of a neon taco truck at night, risograph print texture, only hot pink and teal", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five wrote both lines; Krea 2 and MOSS lost the two-color risograph look; Klein with the upsampler is closest."},
{"id": "cmp-poster-klx", "app": "create-image", "values": {"prompt": "A minimalist concert poster: the title \"MIDNIGHT TACOS\" in huge bold letters at the top, \"LIVE · FRIDAY 9 PM\" in smaller letters below it, an illustration of a neon taco truck at night, risograph print texture, only hot pink and teal", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five wrote both lines; Krea 2 and MOSS lost the two-color risograph look; Klein with the upsampler is closest."},
{"id": "cmp-poster-kr", "app": "styles", "values": {"prompt": "A minimalist concert poster: the title \"MIDNIGHT TACOS\" in huge bold letters at the top, \"LIVE · FRIDAY 9 PM\" in smaller letters below it, an illustration of a neon taco truck at night, risograph print texture, only hot pink and teal", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five wrote both lines; Krea 2 and MOSS lost the two-color risograph look; Klein with the upsampler is closest."},
{"id": "cmp-poster-ms", "app": "moss-image", "values": {"prompt": "A minimalist concert poster: the title \"MIDNIGHT TACOS\" in huge bold letters at the top, \"LIVE · FRIDAY 9 PM\" in smaller letters below it, an illustration of a neon taco truck at night, risograph print texture, only hot pink and teal", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five wrote both lines; Krea 2 and MOSS lost the two-color risograph look; Klein with the upsampler is closest."},
{"id": "cmp-fox-zi", "app": "z-image", "values": {"prompt": "A loose watercolor illustration of a fox reading a book under an oak tree at dusk, visible brush strokes and paper texture, warm and calm", "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Z-Image Turbo (8 steps): the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five fine; MOSS the most saturated."},
{"id": "cmp-fox-kl", "app": "create-image", "values": {"prompt": "A loose watercolor illustration of a fox reading a book under an oak tree at dusk, visible brush strokes and paper texture, warm and calm", "expand": false, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein, prompt as written: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five fine; MOSS the most saturated."},
{"id": "cmp-fox-klx", "app": "create-image", "values": {"prompt": "A loose watercolor illustration of a fox reading a book under an oak tree at dusk, visible brush strokes and paper texture, warm and calm", "expand": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), FLUX 3 Klein with BFL's 9B upsampler: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five fine; MOSS the most saturated."},
{"id": "cmp-fox-kr", "app": "styles", "values": {"prompt": "A loose watercolor illustration of a fox reading a book under an oak tree at dusk, visible brush strokes and paper texture, warm and calm", "aspect": "1:1 (Square)", "enhance": true}, "overrides": [{"node": "30:23", "field": "value", "value": false}],
- "note": "Image model comparison (25 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), Krea 2 Turbo, no style LoRA, its prompt enhancer on: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five fine; MOSS the most saturated."},
{"id": "cmp-fox-ms", "app": "moss-image", "values": {"prompt": "A loose watercolor illustration of a fox reading a book under an oak tree at dusk, visible brush strokes and paper texture, warm and calm", "rewrite": true, "width": 1024, "height": 1024},
- "note": "Image model comparison (25 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image."},
+ "note": "Image model comparison (24 Oct), MOSS-Image SFT with its prompt rewrite: the same prompt on Z-Image Turbo, FLUX 3 Klein (with and without the upsampler), Krea 2 Turbo and MOSS-Image. Verdict (run 20261024-230058-l40s): all five fine; MOSS the most saturated."},
{"id": "demo-music-a", "app": "music", "values": {"style": "instrumental electronic pop, driving synth bass, crisp punchy drums, bright plucked arpeggios, uplifting and modern, steady 120 bpm, no vocals", "lyrics": "", "seconds": 130},
- "note": "Soundtrack candidate for the demo video (made in Freefield, MiniMax Music 3)."},
+ "note": "Soundtrack candidate for the demo video (made in Freefield, MiniMax Music 3). Verdict: a 22 s piece; with empty lyrics MiniMax Music 3 stops early."},
{"id": "demo-music-b", "app": "music", "values": {"style": "instrumental synthwave, pulsing analog bass, gated drums, shimmering pads, confident cinematic build, steady 120 bpm, no vocals", "lyrics": "", "seconds": 130},
- "note": "Soundtrack candidate for the demo video (made in Freefield, MiniMax Music 3)."},
+ "note": "Soundtrack candidate for the demo video (made in Freefield, MiniMax Music 3). Verdict: a 20 s piece, as demo-music-a."},
{"id": "talking-sing", "showcase": true, "app": "talking", "values": {"portrait": "@case:ref-face-a", "audio": "@case:ref-song", "prompt": "She sings along to the song, natural lip movement, subtle head motion, expressive eyes, soft studio light", "duration": 5},
"note": "Baseline audio-to-video: the lips should follow the vocal and the face should stay hers."},
{"id": "lora-talking-head", "app": "talking", "values": {"portrait": "@case:ref-face-a", "audio": "@case:ref-song", "prompt": "She sings along to the song, natural lip movement, subtle head motion, expressive eyes, soft studio light", "duration": 5},