JoyLTX25_Take — one prompt, one take, one generation (8–30 s). A planner sizes the two render passes to your card (x2 / x1.5 / single pass) so it just runs.
JoyLTX25_Multishot — one story, several shots, one video. Each shot is one generation (VRAM = one shot), joined by an AV-extend: the previous shot's last frames and sound are pinned into the next one, so a take can be as long as you like with no visible seam, or you can cut between angles and keep the same voice.
JoyLTX25_Take_PLUS / JoyLTX25_Multishot_PLUS — the same two canvases with four more lanes: a LoRA stack (4 slots), Refs by Name (cast characters from input/joyecho_refs/<name>/ photo folders — one photo per character, and from their second shot on their own rendered face and voice carry into every later shot), Story Source (one panel: your premise, a file of story ideas walked one per queue, or finished shot prompts — with file dropdowns), Refs by Name casting, LoRA stack and audio export, and audio export (flac). Start with the plain canvases; move to PLUS when you want any of those.

Try it without installing anything: the Joy-LTX 2.5 Space (one take, ZeroGPU).

Models: GGUF (RTX 30/40) · comfy-native (RTX 50). Both doses in each: 070T30 (default, cleaner skin) and 100T50 (livelier performances). Pick by card: 16 GB Q4_K_S / w4a8 · 24 GB Q5_K_M / mix4x8-17.0GB · 32 GB Q8_0 / int8 (100T50 = same links with 100T50 in the name; v2 files = LTX-2.5 dev + the JoyAI-Echo delta with the official distilled LoRA baked at 0.5 - the v1 files were over-cooked and are gone). Dev merges (bring your own distill LoRA): https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-dev · Also on Civitai: Joy-LTX 2.5.

Install (5 minutes)

Manager → Install via Git URL → https://github.com/jlucasmcrell/ComfyUI-JoyLTX25 (or unzip into custom_nodes/).
Copy writer_pack/ComfyUI_JoyAI_Echo_GGUF_Nodes/ from this zip into custom_nodes/ (the LLM writer that turns your premise into shot prompts - RealRebelAI's pack with our writing rules bundled; the Manager version lacks them). For .gguf models also install ComfyUI-GGUF.
Put a Joy-LTX DiT file in models/diffusion_models/. 24 GB → …070T30-DiT-Q5_K_M.gguf; 32 GB → …070T30-DiT-comfy-int8.safetensors; 16 GB → …Q4_K_S.gguf or …comfy-w4a8.
From Lightricks/LTX-2.5: the two VAEs → models/vae/, the x2 latent upscaler → models/latent_upscale_models/, the gemma-4 text encoder (…comfy-int8-convrot, or the 10.6 GB w4a8 from LTX-2.5-Quantized for 16 GB) → models/text_encoders/. Optional: ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors from Lightricks/LTX-2.3 → models/latent_upscale_models/ (the planner's x1.5 step for 24 GB cards).
Restart ComfyUI, open workflows/JoyLTX25_Take.json, pick your files in the three loaders, type a premise, Run.

The two canvases

Take: SCENE IDEA → writer (JoyEcho LLMEnhance, hosted model recommended) → JoyLTX Script folds the beats into one LTX prompt → pass 1 at the render size → latent upscale → pass 2 → decode. JoyLTX Take Controls = render size, seconds, upscale. auto reads your VRAM and prints what it planned; x2 / x1.5 / none override it (stream and wait if you like), vram_gb steers auto. USE I2V (off by default): load a picture, flip the gate, and the take starts from that frame (image-to-video) instead of text only. 8 s at 960x544 auto = x2 → 1920x1088 on 24 GB in ~7 min (RTX 3090), ~2 min on an RTX 5090.

Multishot: SCENE IDEA → writer (its num_shots = how many shots) writes N shot prompts → JoyLTX Multishot Sampler renders each shot with the same two-pass pipeline and joins them:

continue — seamless take across generations (writer join style extend take).
cut — only the sound is pinned: the voice carries across a picture cut (writer cuts / continuous scene with cuts).
fresh — independent shots. overlap = pinned latent frames (3 = 0.7 s, trimmed on output). save_every_shot writes each raw shot to output/video/JOYLTX_SHOTS/.

Same people in every shot, without reference photos: identity_ref attaches a frame of shot 1 to every later shot as an in-context keyframe reference (LTX-2.5 AddGuide - it conditions, it never appears), so a cut to a new angle keeps the same face, clothes and room; the voice is carried by the audio pin. identity_strength 0.4-0.7 = same person, free composition; 1.0 = near first-frame copy. Optional shot_images (one per shot) or the USE I2V switch (a Load Image for the first frame of shot 1) if you want to start from your own picture.

Numbers that matter

| card | Take, 8 s 960x544 auto | Multishot, per 8 s shot | |---|---|---| | RTX 5090 (int8) | ~2 min (x2 → 1920x1088) | ~2 min | | RTX 3090 (Q5_K_M) | ~7 min (x2) · ~8 min x1.5 at 1280x736 | ~4 min (x1.5) | | 30 s single pass 1280x736 | 5 min (5090) · 17 min (3090) | — |

Two-pass reads better than a same-token single pass; on 24 GB auto picks x2 for 8 s clips and x1.5 for 12 s at 960x544. Fewer, longer shots mean fewer joins: 4 × 30 s single pass on 24 GB is a 2-minute take with three joins. Cloud writer models sometimes return 503/empty — pick another in the writer's list and rerun.

Nodes

JoyLTX Take Controls · JoyLTX Multishot Sampler · JoyLTX Script · JoyLTX Any Switch (lazy A/B) · JoyLTX Optional Image (the I2V / END-frame gates: emit nothing when off) · JoyLTX Keyframes (end frame and mid keyframes for a take; the sampler has end_images per shot). Plus an in-process patch that fixes ComfyUI's x1.5 latent upsampler when it is only partially loaded (a real crash on 24 GB cards). See SETTINGS.md.

Credits / license

JoyAI-Echo by JD · LTX-2.5 by Lightricks · merge, quants, nodes, canvases by joeygambino. MIT for this pack; the model files follow the LTX-2.x Community License.