Doom LTX Sampler

Doom LTX Sampler is the pack's image-to-video node for LTX-2.5, and it does the whole pipeline in one purple box: first frame + prompt → video with synchronized audio, optional ×2 latent upscale, optional prompt enhancement, and even a duration head that predicts how long the clip should be from your prompt. Feed it the outputs of Doom_LTXLoader, and you get frames, an audio track, and the effective prompt out the other side.

This is the node the pack is quietly built around - the README's LTX workflow is the flagship example. LTX is the speed tier of local video: where Wan takes minutes per clip, LTX takes seconds, and it's one of the few open models generating audio and video together in a single pass.

How it works

The pipeline mirrors the official LTX-2.5 i2v workflow, hardcoded sensibly:

• Two-stage sampling. Stage 1 runs at the base width × height with the distilled 8-step sigma scheme; then, if latent_upscale is on (default), a latent upscaler ×2 brings it to final width×2 × height×2 and a 3-step refine pass runs on top. Turn latent_upscale off for the fast/low-VRAM path - output is then just width × height.
• First-frame conditioning. Your image is the start frame, embedded at the official strengths for each stage. img_compression (default 18, the official value) is the pre-process compression; 0 means none.
• Audio is mandatory-ish. audio_vae is required - sound is generated together with video, not as an afterthought. The audio comes out as audio ({waveform, sample_rate}), ready for Doom_SaveVideoFree or any audio-save node.
• Prompt enhancer. With prompt_enhance on (default), a short prompt is expanded into something cinematic by the instructional Gemma (enhance_clip from the loader), considering the first frame; the enhanced text is returned in enhanced_prompt so you can see what actually went in.
• auto_duration. With the duration head loaded (and auto_duration on), the duration input is ignored and clip length is predicted from the prompt via the official LTX duration predictor, clamped to min_seconds/max_seconds; the raw prediction is emitted as predicted_seconds.
• CFG is fixed at 1.0 - the distilled model's setting, no dial to mis-tune. Sampler default is euler_ancestral.

Outputs: images (frames, straight into a video saver), audio, enhanced_prompt, predicted_seconds. Tiled VAE decoding keeps VRAM sane on lower cards.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

Restart or ComfyUI Manager. No extra pip deps; relies on ComfyUI's own LTX support, so keep ComfyUI current. The README's DoomAI_nodes.git clone line is stale - repo is Doom_Flux_NodePack. All the model files come from the loader: diffusion, Gemma CLIP, video VAE, audio VAE, optional upscaler/enhancer/duration head.

Common issues

• No audio. audio_vae not loaded, or set to None on the loader. It's required here - reconnect it.
• auto_duration on but no duration head → error or ignored toggle. The head must be loaded and connected; otherwise leave auto_duration off and set duration (1–20s) yourself.
• Prompt enhancer doing nothing → enhance_clip missing. Either connect it from the loader or let the sampler's lazy auto-load find the standard Gemma in text_encoders.
• VRAM. Long/high-res clips are heavy. The 22B int8 build is the friendly path; keep latent_upscale off and start at 640×360 if it OOMs. LTX also punishes thin prompts - the KB's LTX doc is full of "seriously low quality" complaints that trace back to under-described scenes.
• img_compression cranked to 0 - sharper start frame, but some workflows behave oddly; 18 is the official value for a reason.

It's the pack's flagship for a reason: video with audio in one node, no sampler constellation, and the distilled defaults already correct.