[rwagwani]

Follow-up to #338, covering the two items it listed as not covered, plus four more Transfer settings found while validating the blur and edge controls of Cosmos3-Nano-Sim-Transfer.

Summary

• A missing negative_prompt_file is now rejected when the sample is parsed, before the model loads. Before, the file was read only when no negative_prompt was set; in almost every setup a default negative_prompt wins first, so a typo in the path was accepted with no message. Lookup order:

  1. the folder the run was launched from (where the file was read from before, so existing setups keep working);
  2. the spec file's folder (specs are validated inside a chdir to it; this is the cookbook's ../assets/negative_prompt.json layout);
  3. for a bare filename only, defaults/ (for example neg_prompts.json).

  A negative_prompt_file inside a --defaults-file is also tried next to that defaults file. A path that is not found fails with:

  negative_prompt_file not found: missing_negative.json (tried <launch>/missing_negative.json, <spec dir>/missing_negative.json, <package>/inference/defaults/missing_negative.json)
  

• negative_metadata_mode and negative_prompt_keep_metadata warn on distilled causal Transfer. They are added to _DISTILLED_TRANSFER_IGNORED_FIELDS, with neutral values none and false, so a request that sets them gets the same [transfer] '<field>' is ignored by distilled causal Transfer: ... warning as the other fix(inference): warn when a transfer setting cannot take effect #338 fields. The distilled schedule has no negative-caption branch, so neither setting is read.

• A preset next to a control_path now warns. load_transfer_control_frames uses a control_path as-is and computes the edge/blur control from vision_path only when there is no control_path, so blur.preset_blur_strength / edge.preset_edge_threshold are never read in that case. A non-default preset set by the request now logs [transfer] 'blur.preset_blur_strength'='very_high' is ignored: 'blur.control_path' is set, .... medium (the default, and the value in the shipped cookbook specs) stays silent.

• show_input without vision_path now warns. There is no input video to put next to the output, so single-view Transfer writes the output alone with no message. A show_input=true set by the request, with every control coming from a control_path and no vision_path, now logs [transfer] 'show_input' has no effect without 'vision_path': ... (streaming Transfer already rejects this request).

• num_conditional_frames >= the chunk size no longer stops the causal path. generate_transfer_sample called _get_num_chunks before branching, so num_conditional_frames=2 with num_video_frames_per_chunk=1 failed after model load with "num_conditional_frames must be smaller than num_video_frames_per_chunk", even though the causal path never reads either value and fix(inference): warn when a transfer setting cannot take effect #338 already reports both as ignored on distilled checkpoints. The chunk partition is now validated only on the chunked (bidirectional) path, which is unchanged.

• resolution given as a number is accepted. "resolution": 480 in a JSON spec failed the Literal check (input_type=int); a before-validator on VisionDataOverrides.resolution turns an integer into its tier string. Values outside the tiers (1080, "4k") and booleans are still rejected as before.

Not changed

• Which negative caption wins when both the defaults and the spec supply one. That changes cookbook output, so it stays a separate decision, as noted in fix(inference): warn when a transfer setting cannot take effect #338.
• Generated output: no caption, sampler or generation path changes.
• Resolution values outside the tiers still get the generic schema message, not the checkpoint's "supports only resolution='480'" message; that needs the check moved after the model config is known, which is a larger change.

Behaviour change, stated plainly

• The only new error is a missing negative_prompt_file, and it lists every path it tried. Every shipped cookbook spec and Sim-Transfer skill request still builds; transfer_shipped_specs_test.py covers this.
• The check runs whether or not the file would be read, including a path in a --defaults-file. A setup that carried an unused, broken path now fails, by design.
• In a batch or jsonl run, a bad path now fails when the samples are loaded, like a bad vision_path, so --skip-invalid-samples does not skip it.
• Requests with num_conditional_frames >= the chunk size on the causal path, and with resolution as an integer, used to fail and now run.
• URLs were never read for this field, and they are now rejected with the same message instead of being accepted unused.
• The stored sample args record the resolved absolute path.
• negative_metadata_mode is checked against the value the request set, so none does not warn when negative_prompt_keep_metadata from the defaults promotes it to same.

Verification

• New L0 CPU tests in sim_transfer_args_test.py:
  • a missing file is rejected on both the causal and the bidirectional path
  • a relative path still resolves from the launch folder first, then from the spec's folder (the cookbook layout)
  • a bare defaults/ filename is still accepted, but a path such as typo_dir/neg_prompts.json is not
  • the keep_metadata promotion does not produce a spurious negative_metadata_mode warning
  • negative_metadata_mode none / same / inverse and negative_prompt_keep_metadata false / true warn only for non-neutral values
• Added for the four new items:
  • sim_transfer_args_test.py: the preset warns for a non-default value with control_path (blur and edge) and stays silent for medium, for no preset, and without control_path; show_input warns only when the request set it and there is no vision_path; integer 480/720 are accepted, 1080, "4k" and True are still rejected.
  • transfer_test.py: the causal path runs with num_conditional_frames 17 and 30 against a 17-frame chunk, for distilled and non-distilled checkpoints; _get_num_chunks still rejects it for the chunked path.
  • Each of these new tests fails without its change (7 failures) and passes with it.
• test_unused_negative_prompt_file_is_not_opened now uses a file that exists, because a missing path is rejected earlier. It still checks that the file is not read when a default caption wins. A stale comment in transfer_shipped_specs_test.py is updated.
• pytest cosmos_framework/inference/sim_transfer_args_test.py cosmos_framework/inference/transfer_shipped_specs_test.py cosmos_framework/inference/args_test.py: 159 passed.
• pytest cosmos_framework/inference/ -m "not GPU": the same 7 failures as main (auto_rig_frontend_test x1, model_test x6), no new ones.
• A spec launched from a folder holding neg.json, with the spec elsewhere, still resolves to the launch folder's file.
• ruff check and ruff format --check: clean.
• After adding the four items and rebasing onto main (d04e506): pytest -m CPU cosmos_framework/data cosmos_framework/inference: 1038 passed, 0 failed; sim_transfer_args_test.py, transfer_test.py, args_test.py, transfer_shipped_specs_test.py, transfer_direct_resize_test.py: 306 passed; pre-commit (base and full config): clean.
• End to end on Cosmos3-Nano-Sim-Transfer (single GPU), with the same spec and the same command:
  • A spec with negative_prompt_file: "missing_negative.json" was accepted silently before. Now it stops before model load with the message above.
  • A spec with a valid negative_prompt_file still runs, with an unchanged result, and logs [transfer] 'negative_prompt_file' is ignored by distilled causal Transfer: the distilled schedule has no negative-caption branch.
  • A spec with negative_metadata_mode: "same" runs, with an unchanged result, and now logs [transfer] 'negative_metadata_mode' is ignored by distilled causal Transfer: the distilled schedule has no negative-caption branch to add metadata to.
  • The four new items, with the blur specs from the parameter sweep (AV clip):
    • the documented blur recipe runs with no new warning;
    • blur.control_path + preset_blur_strength: "very_high" runs and logs the new preset warning;
    • show_input: true with only a control_path runs and logs the new show_input warning;
    • num_conditional_frames: 2 with num_video_frames_per_chunk: 1 failed after model load before; it now runs and logs the two existing "is ignored by distilled causal Transfer" warnings;
    • "resolution": 480 (a number) failed the schema check before; it now runs at 480.

🤖 Generated with Claude Code