LBH 3+1 is a local MiniMax H3 workflow that spends three sampling steps at a smaller video size, enlarges the video latent, and finishes with one higher-resolution step. The ComfyUI-H3-LBH-X2-Stream package combines that schedule with streaming X2 decoding and audio output. It addresses a practical question: can a local creator spend less time sampling while still refining the enlarged result?
This guide reviews the package's public documentation and evidence on October 12, 2026. We have not installed it or generated a video with it. It is a local workflow, not an LBH mode currently available in H3Vid's hosted generator.
Understand the three different resolutions
The selected recipe samples at 384×672 for the first three steps, enlarges the video latent to 544×960, and performs the final H3 step there. X2 decoding then produces 1088×1920 output. Those are three different stages, not three names for native full-HD generation.
Latents are the model's internal video representation. Enlarging them before the final sampling step gives H3 another opportunity to update the image structure. X2 decoding follows afterward. A final file with more pixels does not establish that the subject contains the same detail as a video sampled at that size throughout.
For a first comparison, inspect readable edges, small objects, hair, and facial features. A smooth background is easy to enlarge; a moving hand or patterned garment is a more useful test of whether the finishing step helped. Review the entire motion, not just a sharp still from the result.
Audio must continue across the handoff
The package's handoff implementation and example graph preserve a separate audio path while the video latent changes size. The final H3 step continues both video and audio. Do not send the combined audio-video tensor through a spatial video upscaler or discard the audio branch while rewiring the graph.
Begin with an included workflow and keep its Separate/Concat AV connections intact. If you later simplify the graph, check both sampler outputs and their destinations rather than relying on the visual order of nodes. A graph can produce a playable picture while omitting or mishandling sound.
Use a short scene with a clearly visible sound event for review: a ceramic cup set on a table, for example. Listen to the complete output, then check the contact moment. Dialogue needs another check for voice distortion and mouth timing. These are suggested checks, not results we obtained with the package.
Prepare the local environment before queueing
The author's development record uses Windows, RTX 4070 12GB and approximately 48 GiB of system RAM. The package depends on native H3 support, SLA, Dynamic VRAM, ConvRot W4A4 and Sol-Attn. Having 12GB of GPU memory alone does not establish a matching installation.
Follow the setup implementation with the Python environment that runs your ComfyUI instance. Its installer supports a read-only check, an optional small GPU check, and one-time F12 preparation. It refuses global Python and preserves conflicting existing files. Stop ComfyUI before preparing the local shards.
The setup separates dependency installation from model downloads. Match the filenames in the Japanese model table: the fused INT8 diffusion model, Qwen3-VL NVFP4 encoder, video and audio VAEs, X2 Detail weights, and LBH upscaler all have distinct roles. A similarly named base H3 checkpoint is not automatically the benchmarked model.
Record those selections before changing anything. Save the workflow JSON beside your output, along with the package revision. This makes a later failure or quality difference easier to diagnose than a screenshot of a few nodes.
Start from an included preset
The package includes five-second, ten-second and fourteen-second examples. They use a first image; an FL2VA-derived model does not mean the graph also supplies a final image.
Prepare the X2 decoder with the preparation workflow, then load a video preset and select your own input image. Keep the included sampling schedule for the first run. Change one variable at a time afterward: image, prompt, duration, or a model component. Replacing all of them at once makes the reference evidence much less useful.
Frame counts also explain why the preset label and actual duration differ. At 24fps, 243 frames represent 10.125 seconds and 345 frames represent 14.375 seconds. Use the final file's frame count when aligning music, narration, or an edit timeline.
Read the timing records separately
The historical benchmark record reports 106.879→85.393 seconds for its ten-second comparison and 154.834→116.594 seconds for fourteen seconds. Each method and duration had one measured run. Loading was included; server startup, uploads and validation were excluded. This is evidence about that recipe and environment, not a universal GPU speed claim.
The fresh-install validation is a different record. Its packaged examples completed on ComfyUI 0.39.0, with output decode checks. Passing those checks establishes that files were generated and decoded in the author's environment; it does not establish a quality ranking. The record's process and cache conditions should remain separate from the historical comparison.
A community RTX 5060 Ti report gives another single paired observation, 126.75→105.35 seconds, and explicitly says the comparison video's sound comes from the baseline. It cannot validate the new run's audio by listening to that composite.
Decide whether it fits your next project
Try this route when you already maintain a compatible local H3 stack and want to investigate a specific resolution schedule. Keep a working baseline, label cold and warm runs, and compare several scenes before choosing it as a default. Count rejected clips as well as successful renders when assessing production time.
For general setup, see MiniMax H3 in ComfyUI. For the broader difference between online video enhancement and local latent enlargement, see MiniMax H3 Upscaler. Spark's separate LBH example uses another sampling recipe; the Spark-H3 explanation distinguishes the two.
The cover is an AI-generated concept illustration, not an H3 result, a workflow screenshot, or a benchmark.