ComfyUI workflows, agent access, and official LTX migration

Implemented on 2026-09-22 under WORKSTATION/luciana-media-mcp.

Qwen Image 2.1

Two editable UI workflows and two fixed API graphs use the installed Qwen Image 2.1 INT8 ConvRot transformer, Qwen3-VL INT8 encoder, BF16 VAE, and each final character LoRA. The UI source is the official Comfy-Org template, with a model-only LoRA loader inserted between the transformer and sampler. Defaults are 768×1024, 25 Euler/simple steps, CFG 1, and LoRA strength 1.0. Both profiles generated successfully and were visually inspected. An additional image generated through the actual stdio MCP in 14.257 seconds.

Importable files:

• WORKSTATION/lucianax.json
• WORKSTATION/lucianat.json
• Shared UI library: WORKSTATION/Luciana

The Qwen service is luciana-comfyui.service, loopback port 8191, physical GPU 2. Its shared user directory exposes these workflows in the UI. Custom and API nodes are disabled because these graphs need only core nodes. This also avoids an unrelated custom LTXVideo/Kornia import failure observed during initial startup.

Official LTX-2.5 route

The prior Dev + character LoRA route was a 25-step, low-resolution test. The new route uses the official dedicated Distilled transformer, official Gemma 4 LTX projection, 8+3 manual sigma steps, Euler ancestral, CFG 1/1, x2 latent upscaler, and BF16 diffusion video decoder. The local quality profile changes decode tiling and component placement. These adaptations are recorded in the validation JSON; the upstream template itself is byte-pinned unchanged.

Official source revision: Comfy-Org/workflow_templates@371a7b7171bbd11e9cc92ef615ba5ad223d7e5b4.

• Official I2V
• Official T2V
• Lightricks pipeline reference

All six locked model assets passed SHA-256, size, and SafeTensors structure checks. The missing Distilled transformer and diffusion VAE were downloaded; the encoder was replaced with the exact asset required by the model lock. ComfyUI’s isolated checkout was updated to b0f4b7b2 (0.37.0). The main, dirty ComfyUI checkout was not reset.

The final Dev-trained character LoRA is not used in this official graph. Its compatibility with the dedicated Distilled transformer has not been validated. Character appearance is provided by a Qwen + character LoRA still as the first frame. Consequently these tests assess the official image-to-video path, not whether the trained LTX character LoRA generalizes.

Recorded execution

Both passed full media decode and strict port-8188 history/gallery provenance. Peak sampled VRAM was 23,772 MiB on GPU 0 and 16,928 MiB on GPU 1. All three cards were capped at 245 W. The pair layout places the transformer on GPU 0 and encoder/VAEs on GPU 1. It does not shard the transformer. GPU 2 remains assigned to Qwen.

Evidence: official validation and MCP video validation. Verbatim prompts are omitted from these repository records. Media remains under WORKSTATION/official_ltx25. The turn clip was inspected at its midpoint and showed a coherent profile pose; this is a qualitative smoke test, not a controlled quality or identity benchmark.

Agent integration

The stdio MCP exposes media_readiness, list_character_profiles, generate_character_image, and generate_official_character_video. It returns local media paths. Long calls have a 4200-second client budget, and LTX retains the existing host/worker render locks.

The shared luciana-media skill is symlinked into all five clients. Restart existing sessions to discover new tools. Pi prefixes tool names with luciana_; OMP discovers full schemas through its xd:// tool catalog. The exact same MCP backs the Pi extension.

Reproduction and checks

WORKSTATION/luciana-media readiness
WORKSTATION/luciana-media generate-qwen lucianax 'YOUR DESCRIPTION' --seed 42
hermes mcp test luciana-media

Use generate_official_character_video with the resulting PNG path and a motion description of up to 200 words. It starts the pinned continuity-pair worker if needed, runs the official dispatcher, and requires DONE before returning. For direct CLI usage, follow the installed ltx-25 skill. Official T2V/I2V/FLF2V UI files are installed under the shared LTX-2.5 library; FLF2V is a UI reference and is not exposed by this MCP implementation.

Validation passed: 84 LTX unit tests, static repository provenance, workflow-library integrity, two Qwen graph tests, real image generation for both LoRAs, real MCP image generation, and real official video generation both through the dispatcher and MCP. Dependency-isolated test attempts initially lacked NumPy/pytest; the LTX suite passed with unittest in the existing ComfyUI environment, without adding those packages to the production environment.