IAMCCS-SuperNodes AU+IMG2VID Exec VAE

In the three-stage SuperNode chain (Render → VAE → Finalize), this is the middle step: it takes the latent your Render stage produced and decodes it into actual frames - with an explicit choice of how memory-hungry that decode should be. It can also take a frames_dir directly if you'd rather skip the latent path entirely.

The VAE stage exists because decoding a long video latent is exactly where low-VRAM systems die. LTX 2.x video latents are big, and a naive decode holds everything in memory at once. This node gives you a decode_mode dropdown that trades speed against RAM: normal_tiled_iamccs (default) for normal cards, low_ram / very_low_ram for the rest of us, high_vram when you have headroom and want it fast. The tiled decode options chunk the work, and very_low_ram leans on disk for frame output so system RAM stays sane.

Inputs

• video_latent (optional) - the LATENT from the Render stage.
• vae (optional) - the VAE to decode with.
• frames_dir (optional) - if supplied, the node can work from already-decoded frames instead of a latent.
• decode_mode - normal_tiled_iamccs / low_ram / very_low_ram / high_vram.
• frame_rate - output FPS (default 25).
• filename_prefix, output_root, frames_subdir, image_format (png/jpg), jpg_quality - where and how frames are written.
• tiled_tile_size, tiled_overlap, tiled_temporal_size, tiled_temporal_overlap - tiling geometry; leave alone unless you're tuning low-RAM behavior.
• crf, pix_fmt, trim_to_audio - encode settings used if this stage also produces a video.
• save_metadata, vram_flush, cleanup_before_decode - housekeeping toggles.
• ui_preset - balanced, high_quality, fast_preview, low_ram_safe, very_low_ram_decode. Start here instead of hand-tuning tiles.
• audio (optional) - for trim/encode purposes.
• linx (optional) - pass the SuperNode metadata through.

Outputs

• video_path - a combined video if you let it encode.
• images - the decoded frame batch (careful: this is the big memory object).
• frames_dir_out - where frames were written to disk, ready for the Finalize stage.
• audio_passthrough, linx, report.

Install

Part of IAMCCS-nodes - Manager (search "IAMCCS") or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart. Needs the same base as the other SuperNodes: native LTXV nodes, Video VAE, and ffmpeg.

Practical advice

If you're on a 12GB card and the Render stage succeeds but decode OOMs, you found the reason this node exists - set ui_preset to low_ram_safe or very_low_ram_decode and route through frames_dir_out to disk rather than pulling images into RAM. And if the whole run worked but the final video came out slightly longer/shorter than the audio, verify frame_rate and trim_to_audio here match the Finalize stage - the two nodes won't reconcile each other's numbers.