hermes-skills/ltx-video-pipeline/references/stock-as-i2v-start-frame.md

Stock as I2V Start Frame

Pattern for using stock character images (from TrueNAS ai_vid_stock_material/character_refs/) as direct I2V start frames in the standard LTX pipeline. Simpler than IC-LoRA Ingredients — no additional LoRA, no reference sheet, no static video loop. Works with the existing 6-fix baseline.

When to Use

• Quick character test: "does this stock image animate well?"
• Single-clip renders where character consistency across scenes isn't critical
• Before investing in IC-LoRA Ingredients setup

When NOT to Use

• Multi-scene stories needing character consistency across clips → use IC-LoRA Ingredients
• Stock images with white/light backgrounds → LTX may hallucinate background elements
• Low-resolution stock → resize artifacts will compound in animation

Steps

1. Download stock from TrueNAS:

   smbclient -N //10.0.0.117/proxmoxBackup -c 'cd ai_vid_stock_material\character_refs; get SHEETS2_00005_Boss.png /tmp/boss.png'
   

2. Resize to 768×512 (our standard resolution):

   from PIL import Image
   img = Image.open('/tmp/boss.png')
   img = img.resize((768, 512), Image.LANCZOS)
   img.save('/tmp/boss_start_frame.png')
   

3. Copy to .202 input directory:

   sshpass -p 'passw0rd' scp /tmp/boss_start_frame.png [email protected]:/home/n8n/comfy-ui/input/
   

4. Modify an existing workflow JSON — change the imageFile in timeline_data:

   import json
   with open('/tmp/api_cyberpunk_C1.json') as f:
       d = json.load(f)
   nodes = d['prompt']
   td = json.loads(nodes['131']['inputs']['timeline_data'])
   td['segments'][0]['imageFile'] = 'boss_start_frame.png'
   nodes['131']['inputs']['timeline_data'] = json.dumps(td)
   with open('/tmp/api_boss_test.json', 'w') as f:
       json.dump(d, f)
   

5. Update the prompt to describe the stock character (not the Flux-generated character):

   • Describe what the stock image shows (clothing, features, expression)
   • Keep the same environment/lighting/camera from the original template
   • Target 50-80 words, 2-3 actions

6. Submit and verify (standard pipeline steps 4-8).

Example: Boss Stock

Stock: SHEETS2_00005_Boss.png (3328×2432 RGB PNG, white background) Resized: 768×512 Prompt: "Medium shot, slow steady pan right. A stern boss in a dark suit sits at a terminal in a neon-lit cyberpunk room. Pink and blue holographic light flickers across his face as he types on a glowing keyboard. Over 5 seconds: he leans forward studying the screen with cold intensity, then slowly turns his head right. His chrome temple implant glints in the neon glow. Volumetric fog drifts through colored light from overhead panels. Cables and server racks line the walls, indicator lights blinking. Cinematic lighting, shallow depth of field, 35mm film grain."

Pitfalls

• CRITICAL: Single stock images as I2V start frames do NOT anchor the face. LTX treats the start frame as "general scene composition" not "this exact person." The model will morph the face, change clothing details, and substitute its own training data for the character. Confirmed 2026-07-22: boss stock test — face distorted, did not stay true to image. User: "stock test was malformed. It worked 80% but face distorted and did not stay to image." This approach is useful for quick composition tests only. For character consistency, use IC-LoRA Ingredients.
• Character description in prompt MUST match the canonical descriptions in references/character-descriptions.md. Using generic descriptions like "stern boss" instead of "muscular boss with shaved head and goatee wearing a black suit" causes the model to blend between the stock image and its own training data. The user provided exact descriptions — use them verbatim.
• White/light backgrounds in stock images may cause LTX to hallucinate background elements or wash out the scene. The model was trained on varied backgrounds; a plain white background gives it no environment cues. Prefer stock with scene-appropriate backgrounds, or use Flux to generate a start frame that composites the character into the target environment.
• Stock image aspect ratio rarely matches 768×512. Resize with LANCZOS; avoid stretching (crop to aspect ratio first if needed).
• This is NOT a character-consistency solution across clips. Each clip gets its own start frame; there's no mechanism to keep the same face across scenes. For multi-clip consistency, use IC-LoRA Ingredients.