H3 T2VA tip: reuse generation latents as RefMods to recast characters endlessly
SIR_NVAX_A_LOT · reddit · 2026-10-05
A Redditor shares a workflow for using RefMods in H3 T2VA: instead of feeding 10-20 images to build a RefMod, take the native latent straight from your T2VA character generations.
Key points:
- Once you have a latent you can extend, prepend, or bridge it indefinitely; latents of the same canvas size can be joined
- Latents can be clipped, cropped, scaled, and upscaled for reuse
- Demo shows character Vera recast across multiple T2VA generations at int8/20 steps, no turbo loras
- The author prefers T2VA over R2VA for image quality and plans to try latents from single-frame MiniMax H3 generations
A reusable character-consistency trick.
More from Multimodal
- Opus 5.5 'jawdroppingly great' at video, early hands-on praise emerges — vivekhaldar · 2026-10-05
- ACE-Step 1.0's weak audio quality fixed by extracting melodies with SheetSage2 and remixing in YuE2 — Fun_Firefighter_7785 · 2026-10-05
- Free FLUX2 x LTX2.5 ComfyUI workflow series for AI architectural imagery — paulhax · 2026-10-05
- 400 images later, a DGX Spark user remains unconvinced by SenseNova U1.5-8B — dh7net · 2026-10-05
- Citrine Photo: free open-source Windows app runs SeedVR2 upscaling without ComfyUI — garionhk · 2026-10-05
- Grok's image generation quality keeps improving, just in time for Halloween — yunta_tsai · 2026-10-05