LTX-2.5 22B Workflow Chains 6 Keyframes into 50s of Video on Just 8GB VRAM
Ecstatic-Use-1353 · reddit · 2026-09-07
Reddit user reventadirecta released an open-source ComfyUI workflow that pushes LTX-2.5 22B to generate long, controllable video sequences on low-VRAM hardware.
- Core idea: instead of a single first/last-frame pair, chain 6 keyframes with 5 independent prompts across 5 sequential FLF2V generations, letting scene, camera and action evolve progressively
- Tested on RTX 4060 Laptop (8GB VRAM) + 16GB RAM, ComfyUI 0.34.3, Windows
- Stack: LTX-2.5 22B Distilled, W4A8 ConvRot transformer, Gemma 12B W4A8 text encoder, BF16 video/audio VAE, DynamicVRAM enabled
- Results: 720×1280 full pipeline ran in 30m54s producing 50s of 24fps directed video; 576×1024 recommended for iteration
- Sequential execution via ImpactExecutionOrderController keeps peak VRAM manageable; a custom LTX25VideoConcat node stitches clips (H.264, CRF 18, no audio mux)
Repo includes workflow JSON, custom node, deps, model list, and bilingual docs: github.com/reventadirecta/LTX2.56FRAMES5PROMPTS
More from Multimodal
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26