Doom LTX Loader

Doom LTX Loader is the pack's single-node front door to LTX-2.5 image-to-video. LTX's video stack is not one model - it's a diffusion transformer, a Gemma 4 text encoder, a video VAE, a separate audio VAE, an optional latent upscaler, an optional prompt-enhancer Gemma, and (newer) a duration-prediction head. Stock ComfyUI wants a small constellation of loader nodes for that. This loader pulls the whole constellation into one node and hands the sampler everything it needs through dedicated outputs.

LTX is the speed tier of local video - the KB's panel frames it as "seconds-long drafts on consumer VRAM" against Wan's minutes - and LTX-2.5 is the pack's target: a 22B distilled model shipping as the int8-convrot build, which matters because int8-convrot is the quantization path that works on 20/30/40/50-series cards (the ones without fp8 support). If you've been running video at all in 2026, this is the "let it actually run on my card" story.

What it loads

The inputs are all file pickers with clear tooltips naming the exact files:

• model_name - the diffusion model, e.g. ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors in diffusion_models.
• clip_name - the Gemma 4 encoder with LTX projection, loaded with CLIPType.LTXV.
• vae_name - the video VAE (ltx-2.5-video-vae-bf16.safetensors); the loader determines the VAE type and config from the file's metadata automatically, same as stock VAELoader.
• audio_vae_name - the audio VAE (ltx-2.5-audio-vae-bf16.safetensors), None if you want no sound. Note the sampler treats audio as required once you're making audio-video; see below.
• upscale_model_name - the latent spatial upscaler ×2, optional. Powers the sampler's latent_upscale path.
• enhance_clip_name - a lightweight instructional Gemma (gemma4_e2b_it_bf16) used by the sampler's prompt enhancer. Defaults to the standard file if it's present in text_encoders.
• duration_head_name - the duration head (model_patches/ltx-2.5-duration-head-bf16.safetensors), optional, enabling the sampler's auto_duration mode where clip length is predicted from the prompt rather than hand-set.
• weight_dtype - default / fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2 for the diffusion weights.

Outputs: model, clip, vae, audio_vae, upscale_model (LATENT_UPSCALE_MODEL), enhance_clip, and duration_head (MODEL_PATCH). Wire them into Doom_LTXSampler in one pass.

Installing it

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

Restart or ComfyUI Manager. No extra pip deps - it leans on ComfyUI's own LTX support (comfy_extras), so keep ComfyUI current. README's DoomAI_nodes.git line is stale; the live repo is Doom_Flux_NodePack. Model files go in diffusion_models, text_encoders, vae, latent_upscale_models, and model_patches respectively.

Common issues

• Everything but the download is easy. This loader covers a lot of files, and the first failure mode is simply missing ones - check that each filename exists in the folder the tooltip names before suspecting the node.
• Audio VAE errors usually mean you selected a file whose metadata doesn't identify it as the audio VAE - the loader validates (throw_exception_if_invalid). The corrected/current LTX audio VAEs have the metadata; old renamed copies may not.
• enhance_clip default surprises people - the loader pre-fills the standard instructional Gemma when the file exists, so don't be puzzled by a non-None default; it's intentional.
• The Pro/Free split. The pack's 2.0.0 changelog notes the extended LTX sampler (video_cfg/audio_cfg, refine_strength) moved to the paid DoomAI_Pro pack. This free loader/sampler pair is the clean distilled path: fixed CFG 1.0.

If your goal is "LTX-2.5 video with sound, without a six-node loader graph," this is the loader that gets you to the sampler with everything in hand.