[Recipe] Add Cosmos3-Edge

[lishunyang12]

Summary

#5001 added support for the Cosmos3 Edge checkpoint; this adds the recipe for it. recipes/cosmos3/ covered only Nano and Super.

Edge is worth its own page rather than a note on an existing one, because its defaults differ enough that copying Nano's or Super's parameters lands on the wrong operating point:

	Edge	Nano / Super
T2I resolution	640×640	1024²
T2V / I2V resolution	480×832	1280×720
Video guidance_scale	5.0	6.0
Video flow_shift	3.0	10.0

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[lishunyang12]

@bastefaniak PTAL cc @MaciejBalaNV

[MaciejBalaNV]

Action is supported as well

[lishunyang12]

Noted.

[MaciejBalaNV]

Why only B300?

[lishunyang12]

Thanks for reminder, i don't think B300 recipe examplifies downstream usage. I will remove it, and leave follow-up prs to provide benchmarking results.

[MaciejBalaNV]

These generation number look way too high

[MaciejBalaNV]

Can we add the mention about action generation? Other than that, LGTM

[lishunyang12]

Can we add the mention about action generation? Other than that, LGTM

Done

[lishunyang12]

Self-review notes after checking this page's claims against main (post-#5001). Flagging what I found so it's visible to reviewers; I'll push fixes.

Verified accurate: defaults table matches the constants exactly (pipeline_cosmos3.py:173-178 — 640×640 T2I, 480×832 T2V, guidance 5.0, flow_shift 3.0), including the I2V path (:491-495); backbone auto-selection works as described (:232-237, nothing to pass); the three rejections are real (_validate_edge_generation_mode, :1751-1765); both guardrail mechanisms are correct (serve.py:741 for --no-guardrails, text_to_image.py:447 for the "guardrails" extra-body key).

1. The acceleration note is wrong (Cosmos3-Edge.md, Support notes)

Cache-DiT is a silent no-op (no _cache_dit_adapter_config); Sequence Parallel is unvalidated (no _sp_plan)

Both parentheticals are false. Cosmos3EdgeVFMTransformer subclasses Cosmos3VFMTransformer (transformer_cosmos3_edge.py:303), which defines _cache_dit_adapter_config at transformer_cosmos3.py:989 and _sp_plan at :1020. Edge overrides neither. The adapter config targets gen_layers, which Edge has (validate_loaded_weights requires gen_layers.{i}.mlp.*), and the SP plan targets the GEN pathway Edge shares.

Cache-DiT is therefore wired on Edge, not a no-op — as written the page tells users to skip a feature that works. Will re-test and rewrite this section based on measured behavior rather than attribute presence.

2. The ffprobe expected output can't occur (Verification)

The default writer emits [STREAM] / key=value lines, so video,832,480,49 never prints. The field order is also flipped — the command requests codec_type,nb_frames,width,height, the expected line reads codec_type, width, height, nb_frames. With -of csv=p=0 it would be video,49,832,480.

3. Every tuned parameter in the examples restates a default

T2I passes --num-inference-steps 50 --guidance-scale 7.0 + flow_shift: 3.0; constants are 50 / 7.0 / 3.0 (:168-170). T2V passes 35 / 5.0 / 3.0; constants are 35 / 5.0 / 3.0 (:161,175,176). Only --num-frames 49 and max_sequence_length are non-default. The page's own advice is to omit these to get the Edge defaults, so the examples work against it — and the values will drift silently when the constants change.

4. 1x GPU is not a useful hardware cell (recipes/README.md)

Every other row names a part and VRAM. For the checkpoint whose pitch is constrained deployment, that column is the number readers come for. Will fill in the GPU actually tested.

Minor

• nvidia/Cosmos3-Edge-Policy-DROID — needs confirming, I haven't verified this ID resolves.
• Ordering: Edge sits between Nano and Super in the table, but the page describes it as the smallest checkpoint.

[david6666666]

LGTM

[princepride]

lgtm