Skip to content
Python

docs: runs 09-11 in the run log; README totals (31 runs, 591 of 614 cases, 262 cases)

neuregexa37db8d2a3b0
diff --git a/README.md b/README.md
index 39b3b3f..835fd4a 100644
--- a/README.md
+++ b/README.md
@@ -22,9 +22,9 @@ 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 |
| 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 | 28 lab runs on Modal from 20 to 24 Oct (L40S and H100), 581 of 604 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 | 31 lab runs on Modal from 20 to 24 Oct (L40S and H100), 591 of 614 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 | 261 cases: one showcase per app (110), one per effect (24), one per look of `mixed-media` (16) and one per transformation of `morph` (15), 33 references the others start from, 6 candidate portraits for those references (`refx-*`), 12 style-drift experiments (`mmx-*`) and 36 template experiments (`tplx-*`), each experiment with its verdict in the note |
+| LoRA lab | 262 cases: one showcase per app (110), one per effect (24), one per look of `mixed-media` (16) and one per transformation of `morph` (15), 34 references the others start from, 6 candidate portraits for those references (`refx-*`), 12 style-drift experiments (`mmx-*`) and 36 template experiments (`tplx-*`), each experiment with 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
@@ -245,7 +245,7 @@ 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 (137 cases)
+modal run --detach lab/lab_modal.py --showcase # one example per app + the references they need (138 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 bf992fa..8a2c84e 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. 28 runs, 581 of 604 cases ok: 22 of the 23 failures are
+steps loaded, so a peak there is not the least an app needs. 31 runs, 591 of 614 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 |
@@ -36,6 +36,9 @@ the first IC-LoRA run, before the gated licenses were accepted, and the other is
| `20261024-164909-h100` | H100 | Queue job 06: the three new woman references and the 40 examples made from them | 48 / 48 | 215 s | 50.0 GB |
| `20261024-172353-h100` | 2 × H100 | Queue job 07: the ten new apps, `cinema-studio` and `video-t2v` | 20 / 20 | 204 s | 50.5 GB |
| `20261024-174729-h100` | H100 | Queue job 08: `expand-image` on a blurred canvas, `cinema-studio` with looks instead of gear | 3 / 3 | 25 s | 13.2 GB |
+| `20261024-180125-h100` | H100 | Queue job 09: the runner reference with a named person, and the two cases that read it | 3 / 3 | 197 s | 39.9 GB |
+| `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 |
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.
@@ -64,7 +67,10 @@ one folder. The cases a run lists include the reference images, clips and songs
- **References of people.** Z-Image tends to draw an East Asian face when a prompt names no one, and those portraits
looked synthetic. On 24 Oct the three woman references (`ref-face-a`, `ref-fx-woman-a`, `ref-fx-woman-b`) got
prompts that say who the person is, plus natural skin texture and no retouching; job 05 tried two candidates for
- each (`refx-*`, the verdict in each note) and job 06 re-made the 40 examples that start from them.
+ each (`refx-*`, the verdict in each note) and job 06 re-made the 40 examples that start from them. The runner
+ (`ref-fx-runner`) got the same change in jobs 09 to 11. Z-Image also printed a sports brand logo on her top, in a
+ red and white kit and again in a teal one, whatever the prompt said, so Klein removes it (`ref-fx-runner-plain`)
+ before Bullet Time Splash and the Hero run effect use the photo.
- **New apps.** Job 07 made the first examples of `depth-map`, `video-remove-bg`, `h3-video`, `h3-photo`,
`extend-video`, `expand-image`, `ai-influencer` and the three Bullet Time apps. Two came back wrong and were fixed in
job 08: `expand-image` drew a gray frame on the flat padding (it now starts from a blurred, stretched copy of the