TeleStyle Video Transfer

This is the main event. TeleStyle Video Transfer takes the model object from TeleStyle Model Loader, your input frames, and an optional style reference image, then runs the actual stylization: a denoising pass later you get a stylized version of your clip back as an IMAGE batch. The loader is the engine; this node is the gas pedal.

What it's doing under the hood

The source video and the style reference both get encoded into Wan's latent space by the 3D causal VAE. The reference image becomes the "first frame" conditioning latent - that's the TeleStyle reference-to-video trick, the same family as Wan VACE, tuned for style instead of identity. The actual style itself comes from the pre-computed prompt_embeds.pth the loader baked in; there is no text encoder anywhere in this loop, which is why the workflow is so short.

Two quiet things happen before denoising: your frames get padded to the 4n+1 count Wan's temporal VAE requires (by repeating the last frame), then cropped to multiples of 16 on height and width. Both get undone on the way out, so you get back exactly as many frames as you fed in.

Then it denoises on the Wan 2.1 1.3B transformer with a flow-matching scheduler. At the default cfg of 1 there's no classifier-free-guidance pass at all - the whole "negative prompt" idea is skipped, which is standard for flow-matching models. Raise cfg above 1 and you get the second, unconditional denoise pass and real CFG steering.

The inputs a beginner actually sets

• video_frames - any IMAGE batch: VHS Load Video, VideoHelperSuite's Load Video, even a Load Image. Frame count of your input = frame count of your output.
• style_image (optional) - a single image that steers the style. This is the trap: leave it empty and the node feeds random noise as the reference first frame, so the baked-in style embedding has to do all the work and results get unpredictable. Feed it a reference.
• steps (default 12), cfg (default 1), seed (default 42) - the usual dials. 12 steps is already short for a flow model; going lower is for drafts.
• scheduler - FlowMatchEuler, UniPC, or DPM++ (default). All three are adapted from the same flow config with Wan's shift=3.0; DPM++ tends to look best per step here.
• fast_mode (default True) - keeps the whole pipeline on the GPU. Flip it off for sequential CPU offload; slower, but it's how you fit this on 6-8 GB cards.
• enable_tiling (default False) - VAE tiling for big clips. It also switches itself on automatically when you feed more than 16 frames.
• acceleration - default / flash_attn / sage_attn / mem_efficient. Read the source, not the README, on this one. The README advertises Flash Attention 2 and SageAttention support and suggests installing sageattention manually, but the shipped code never imports either library. All four options are just PyTorch SDPA backend toggles (torch.backends.cuda.sdp_kernel), and on recent PyTorch that deprecated API can silently no-op back to the default path. Leave it on "default"; installing sageattention will not speed this node up.

Wiring it up

The reference workflow is tiny: VHS Load Video → TeleStyle Video Transfer → VHS Video Combine, with a LoadImage → resize chain feeding style_image. The output is a 0-1 float IMAGE batch, same frame count as the input, ready for VHS_VideoCombine or VideoHelperSuite's composite to write out.

Set expectations

TeleStyle the research landed in January 2026 and the crowd was cautiously impressed, with two recurring complaints: the samples were "really bent on not turning their heads," and output could look "not fully resolving." This port runs the same idea on a 1.3B model - think quick stylization drafts and mood tests, not your final render.

Troubleshooting

• Out of VRAM → fast_mode off, enable_tiling on, shorter clip, or rebuild the loader in fp16.
• Style barely shows → you left style_image empty. Feed a reference.
• First run "hang" → that's the loader downloading ~6 GB, not this node.
• Long clips degrade → past 16 frames tiling kicks in automatically and consistency fades the longer you go; the pack's own to-do list admits "consistency for very long videos" is unfinished.