Slice 6: video LTX-Video 2.3 22B t2v + i2v (#8)

[jmpijll]

Closes #8

Summary

Adds two new video-Modality Manifests for LTX-Video 2.3 22B - text-to-video and image-to-video - and chains the i2v variant from qwen_image_2512 so a finished Qwen image can be animated end-to-end. Reuses the slice 5 VideoPlugin verbatim (no plugin changes).

Acceptance criteria

• workflows/ltxv_2_3_22b_t2v.json - text-to-video graph
• workflows/ltxv_2_3_22b_i2v.json - image-to-video graph
• workflows/manifests/ltxv_2_3_22b_t2v.yaml (slots: prompt, negative_prompt, width, height, frame_count, seed, lora, lora_strength)
• workflows/manifests/ltxv_2_3_22b_i2v.yaml (above + init_image at attachment_position 1; chained from qwen_image_2512's output_image)
• Reuse VideoPlugin - no plugin changes (the slice 5 mapper handles LTX's single SamplerCustomAdvanced Progress stream as the degenerate case of its WAN 2.2 dual-stream design)
• Live smoke for both manifests against http://172.27.1.165:8188 - both SMOKE OK
• i2v smoke source image generated by qwen_image_2512 first
• Unit tests: tests/test_ltxv_manifest_loading.py + tests/test_video_plugin_ltx_compat.py
• pytest passes locally (369 passed, 1 pre-existing skip)
• All filenames in requires: resolved live against /object_info

Live smoke evidence

Full transcripts + reasoning in docs/v3/smoke/SLICE_6_RUNS.md.

Manifest	prompt_id	size	latency
ltxv_2_3_22b_t2v	5955945b-0162-4730-b886-99d82c1a42ca	384 KB MP4	33.04 s
qwen_image_2512 (source for i2v)	402f9033-617c-4413-8d2d-b39128c974ec	3.13 MB PNG	27.71 s
ltxv_2_3_22b_i2v	c043322d-886f-44da-902c-ce978248eb67	459 KB MP4	42.85 s

Plugin change summary

No plugin changes - VideoPlugin handled LTX-Video 2.3 22B as-is.

• The dual-sampler progress mapper sums per-node Progress streams; with one stream it degrades to a clean monotone 0->95% then bumps on post-sample Executing (VAEDecode, VHS_VideoCombine) and flips to 100% on ExecutionComplete. Covered by TestSingleSamplerProgressMapping in tests/test_video_plugin_ltx_compat.py.
• The MP4 render_outputs path and validate_slot_values coercion both work as-is against the LTX manifests' slot shapes (width/height multiple_of=32, frame_count min=9 max=257, dynamic LoRA enum from comfyui.loras).

Server-side notes for follow-up

During live discovery two nodes the slice 6 ticket suggested ("at minimum") turned out to be unusable on this server install (not a workflow bug):

1. LTXVGemmaCLIPModelLoader needs a Gemma tokenizer.model sentencepiece file that isn't installed. Used LTXAVTextEncoderLoader (LTX 2.x successor, loads the same Gemma weights) instead.
2. LTXVQ8LoraModelLoader imports hadamard_transform which isn't installed. Used plain LoraLoaderModelOnly instead - the LTX distilled LoRA is in its option list.
3. LTXVGemmaEnhancePrompt raised min() iterable argument is empty against the LTXAVTextEncoderLoader CLIP. Skipped; the user's prompt feeds straight into CLIPTextEncode.

These are server install gaps, not architectural choices - documented in docs/v3/smoke/SLICE_6_RUNS.md so a follow-up slice can re-enable the LTX-specific loaders once the missing files / Python deps land.

Anchors

• PRD: §3 Modality coverage (video), §5 Manifests as data
• ADR-0001 (manifest format) - new manifests obey the schema verbatim
• ADR-0002 (modality plugins) - VideoPlugin reused verbatim
• ADR-0003 (slot UI hints) - dynamic LoRA enum via options_from: comfyui.loras

Test plan

• pytest -q -> 369 passed, 1 skipped
• pytest tests/test_ltxv_manifest_loading.py tests/test_video_plugin_ltx_compat.py -q -> 43 passed
• Live t2v smoke against ComfyUI -> SMOKE OK
• Live i2v smoke against ComfyUI with init_image uploaded via _upload_image_slots -> SMOKE OK

Process

Ran from worktree /Users/jamievanderpijll/discomfy-slice-6-ltx (per the slice 6 issue's "workspace isolation" rule); no files modified in the parent /Users/jamievanderpijll/discomfy checkout.

Made with Cursor