DucHaiten-Highway V1

Video, stereo audio and multimodal references — together in one model for ComfyUI.

15-second reference-guided scenes · Four-image demonstrations · Direct high-resolution generation · Standard Safetensors checkpoint · Local RTX 4090 execution

Highway V1 brings its model definition, transformer payload, generation policies and resource management together in a loadable checkpoint with a matching Highway runtime. It supports creative workflows from a short text prompt to multi-reference action scenes, with audio generated alongside the video.

Highway V1 compared with MiniMax H3

20 denoiser evaluations. Quality matching the 50-step H3 results in my evaluated comparisons.

Comparison	MiniMax H3	Highway V1 INT8
Quality in my evaluated comparisons	50-step reference	Matched at 20 denoiser evaluations
Sampling control	Documented 20- and 50-step INT8 recipes	Native C20, saved with the model; no Steps slider
Local output resolution	Released base recipe targets a 768-pixel short edge	1344×768, 1792×1024 and 2688×1536 demonstrated on RTX 4090
Checkpoint coverage	Separate FL2VA and Ref2VA partitions for text/first-last and multimodal references	One Highway checkpoint exercised across text-only, four-image references and first/last-frame generation
Video and audio	Joint video/audio generation	Video with 32 kHz stereo audio, including a 15.08-second, four-reference demonstration

In my evaluated comparisons, Highway’s native C20 schedule (20 evaluations) matched the quality of MiniMax H3 at 50 steps. The quality row records my visual assessment; the resolution and input rows compare the published H3 recipe with the Highway demonstrations below. They are not a same-input speed or memory-saving benchmark. The published 768p target is not an OOM claim or a proven upper limit for H3. H3 source documentation · Measurement details.

Five measured runs on RTX 4090

Demonstration	Frames / playback fps	Complete request	Worker RAM peak (RSS)	Worker GPU allocator peak (allocated / reserved)
PUB-01 · Text only · 1344×768 · 5.17 s	124 / 24	359.34 s	29.00 GiB	3.31 / 3.78 GiB
PUB-02 · Four references · 1344×768 · 5.17 s	124 / 24	524.24 s	29.72 GiB	3.81 / 4.41 GiB
PUB-03 · Four references · 1344×768 · 15.08 s	362 / 24	1,473.58 s	32.50 GiB	7.36 / 9.72 GiB
PUB-04 · First/last images · 1792×1024 · 5.17 s	124 / 24	671.34 s	30.83 GiB	4.97 / 6.62 GiB
PUB-05 · Text only · 2688×1536 · 5.17 s	124 / 24	1,972.19 s	33.35 GiB	9.03 / 11.97 GiB

Memory scope: RAM tracks recorded Highway processes, excluding unrelated applications. GPU figures are the generation worker's PyTorch allocator counters; they exclude the separate image reader and memory outside that allocator, so they are not total Highway VRAM. Reserved includes allocated. These are measured peaks, not minimum hardware requirements. The complete-request times include preparation and output work on a shared workstation; no background time was subtracted. The historical build identities and exact boundaries are listed in BENCHMARKS.

Demonstrated results

Achievement	What Highway V1 has demonstrated
Longer scenes with multiple references	362 frames at 24 fps — 15.08 seconds — at 1344×768, conditioned on four images and accompanied by stereo audio.
Short-form local generation	A 1344×768, 124-frame text-only video: 238.06 seconds for generation; 359.34 seconds for the complete request on the measured RTX 4090 setup.
Measured process memory	For the short text-only case: 29.00 GiB worker RSS and 3.31 GiB allocated / 3.78 GiB reserved in the generation worker's GPU allocator. Unrelated applications are excluded; the allocator figure is not total Highway VRAM.
Visual control	Four-image reference demonstrations at 1344×768, and first/last-frame generation at 1792×1024, with distinct opening and closing actions.
Direct high-resolution generation	A 2688×1536, 124-frame showcase generated at the target dimensions, without first producing a completed smaller video.
Audio included	The measured outputs contain 32 kHz stereo audio together with the video.
An independently loadable transformer artifact	Highway loads its own transformer payload without reopening the source H3 transformer checkpoint. The artifact is approximately 20.97 GB; encoder/VAE and runtime dependencies remain separate.
Architecture work without conventional retraining	The V1 work integrates execution and conditioning around inherited learned weights, without conventional retraining on a new video dataset.

These are measured demonstrations, not performance projections. All five videos passed technical checks and my qualitative review. Timings and memory are single-run observations on a shared workstation; the reference samples used separately identified reader candidates. Full results, build identities and measurement boundaries.

Get started with ComfyUI

Downloads and requirements · Installation guide · Workflow bundle · My Hugging Face profile

On Windows: download the ComfyUI integration ZIP, extract it, close ComfyUI and double-click INSTALL_WINDOWS.bat. Select your ComfyUI folder if asked, then approve the displayed software downloads. The installer handles Python, libraries and FFmpeg without changing ComfyUI's own environment.

Restart ComfyUI, drag in 01 - Start Here.json, and click Prepare / download models on Asset Profile. The panel finds compatible files already on disk, downloads missing files with pause/resume and SHA-256 checks, and saves the paths. Required model files total 78.30 GB on an empty installation; the optional reference reader adds 8.89 GB. Gated downloads still require your own account access.

The ZIP includes all three workflows. The checkpoint is prepared through the model panel; you do not need to download each dependency or edit JSON yourself. No separate Highway application is required. Installation guide · Model setup.

Why is there no Steps slider?

Highway manages its generation schedule; you do not need to choose a sampling-step count. This is an intentional part of the model interface, not a missing ComfyUI control.

The current native generation path performs 20 denoiser evaluations using the schedule and update policy saved with the model. Generation still proceeds through internal numerical updates. The evaluation count is fixed in this version rather than adjusted automatically to scene content.

Setting an ordinary sampler to 20 steps does not reproduce Highway's execution policy. Where evaluations occur and how the state is updated matter alongside their count. Keeping these decisions together preserves the versioned behavior that was tested. Video/audio shift and editing-strength controls retain their own meanings; they are not substitutes for a Steps slider.

In my evaluated comparisons, Highway’s native C20 schedule (20 evaluations) matched the quality of MiniMax H3 at 50 steps. All five demonstrations below use Highway’s native schedule and passed my quality review.

Watch the demonstrations

I used quickly drafted test prompts without extensive tuning. The examples show the resulting outputs as generated. Prompt limitations do not establish the cause of every visible issue.

Each player retains the original video and stereo audio. Press play and enable sound. The prompt panels contain the exact submitted user text, with line endings normalized for display; automatic reference assistance is identified separately.

Measurement key: one observed run per case on a shared RTX 4090 workstation. RAM is measured for recorded Highway processes; GPU allocated/reserved are generation-worker allocator counters, excluding the separate reader and allocations outside that allocator. Unrelated applications are excluded from these memory figures. Reserved includes allocated; do not add them. No whole-device VRAM, power, energy or temperature figure is presented as a Highway-only cost.

Quick comparison and full measurement methodology

PUB-01 · Text-to-video with stereo audio

1344×768 · 5.17 seconds · 24 fps · stereo audio

Open original MP4

Measurement / setting	Recorded value
Output	1344×768 · 124 frames · 24 fps · 5.167 s
Audio	32 kHz stereo
Seed / video shift / audio shift	24681 / 12 / 5
Generation / complete request	238.06 s / 359.34 s
Worker RAM peak (RSS)	29.00 GiB
Owned process-tree RAM peak (RSS / USS / private commit)	29.81 / 24.31 / 35.59 GiB
Generation-worker GPU allocator peak (allocated / reserved; reader excluded)	3.31 / 3.78 GiB
Measured build	rc2-original

No reference images. The output includes physical arrow/quiver elements despite the energy-only prompt.

Prompt — read the exact submitted text

Goal: Show a brief, continuous fantasy fight between two adult opponents in a weathered stone courtyard above a wooded gorge. Neith is a female archer with short dark braids, a copper-brown leather jacket, dark trousers and a solid golden bow. Elec is a male fighter with close-cropped silver hair, a charcoal coat, sturdy boots and a wooden cane. Exactly these two fighters remain in view. Stages: 0-2 seconds: Elec advances with a sweeping cane strike. Neith sidesteps and draws her bow while moving; an arrow made entirely of golden light forms between her hand and the bow. 2-3.5 seconds: Neith releases the energy arrow. Elec deflects it with his cane; brief golden sparks mark the contact, then fade. 3.5-5 seconds: Elec closes the gap and swings again. Neith pivots, intercepts the cane with her solid bow and continues moving as the clip ends. Show clear approach, contact and follow-through. Camera: One continuous medium-wide tracking shot with a restrained arc, both fighters and their footing visible. Cinematic 3D realism, natural daylight, coherent weight and ground contact. The courtyard and distant gorge stay fixed in the world as perspective changes. Audio: Synchronized footsteps, moving cloth, cane-and-bow contact, a short energy draw, release and impact sound, and quiet wind. No dialogue, narration or music.

Exclusions: No extra people, physical arrow shafts, arrowheads, feathers, quivers, glowing eyes, costume swaps, slow motion, idle aiming, teleportation, warped architecture, scene cuts, labels, subtitles, logos or black borders.

Plain-text prompt

Reference images: none — text-only generation.

PUB-02 · Four references in one scene

1344×768 · 5.17 seconds · 24 fps · stereo audio

Open original MP4

Measurement / setting	Recorded value
Output	1344×768 · 124 frames · 24 fps · 5.167 s
Audio	32 kHz stereo
Seed / video shift / audio shift	24681 / 12 / 5
Generation / complete request	275.60 s / 524.24 s
Worker RAM peak (RSS)	29.72 GiB
Owned process-tree RAM peak (RSS / USS / private commit)	31.11 / 25.27 / 37.03 GiB
Generation-worker GPU allocator peak (allocated / reserved; reader excluded)	3.81 / 4.41 GiB
Measured build	rc2-readerfix104-r2-experimental

All four images below were supplied. Automatic assistance added two character associations and limited clothing cues; environment and bow associations were not confirmed separately.

Prompt — read the exact submitted text

Goal: Show Neith, the female archer, and Elec, the male opponent, exchanging attacks continuously for five seconds. Neith keeps her solid golden bow and Elec keeps his cane. Exactly two fighters, coherent positions and momentum. Stages: 0-2 seconds: Elec advances with a sweeping cane strike. Neith sidesteps and draws the bow while moving; an arrow made entirely of golden light forms as she draws. 2-3.5 seconds: Neith releases toward Elec. He deflects the energy arrow with his cane, creating a brief localized burst of golden sparks, and continues closing the gap. 3.5-5 seconds: Elec follows through into another strike. Neith turns outside its line and intercepts the cane with the solid bow. Their exchange remains active through the final frame. Camera: One continuous tracking shot with a restrained arc. Keep both fighters, their footing and the path of each attack readable. Preserve coherent perspective and background parallax during camera movement, with natural ground contact and material response. Audio: Synchronized footsteps, cloth, cane-and-bow impacts, short golden-energy formation, release and impact sounds, and quiet environmental wind. No dialogue, narration or music.

Exclusions: No extra characters, physical arrows, quivers, glowing eyes, equipment swaps, slow motion, frozen aiming, teleportation, scene cuts, distorted architecture, contact sheets, labels, subtitles, logos or black borders.

Plain-text prompt

Actual input images — select an image to inspect the original

PUB-03 · A continuous 15-second reference-guided scene

1344×768 · 15.08 seconds · 24 fps · stereo audio

Open original MP4

Measurement / setting	Recorded value
Output	1344×768 · 362 frames · 24 fps · 15.083 s
Audio	32 kHz stereo
Seed / video shift / audio shift	24681 / 12 / 5
Generation / complete request	1,164.05 s / 1,473.58 s
Worker RAM peak (RSS)	32.50 GiB
Owned process-tree RAM peak (RSS / USS / private commit)	34.00 / 28.14 / 42.48 GiB
Generation-worker GPU allocator peak (allocated / reserved; reader excluded)	7.36 / 9.72 GiB
Measured build	rc2-readerfix104-experimental

The same four visual references were supplied. This reader candidate added two character associations without the extra clothing cues used in PUB-02; this is not a duration-only comparison.

Prompt — read the exact submitted text

Goal: Create a 15-second cinematic fantasy fight between Neith, the female archer, and Elec, the male opponent. The fight is already in progress on the first frame and continues actively through the final frame. Use one continuous shot with coherent movement and clearly readable attacks, evasions, and contact. Continuity: Keep exactly two distinct fighters throughout the fight. Neith keeps the bow in her hand as she moves; Elec keeps his cane. Maintain their positions, momentum, distance, and orientation from one action into the next. Every arrow Neith fires is made ENTIRELY of golden light energy. It forms as she draws the bowstring, leaves the bow on release, travels continuously toward its target, and breaks into brief fading light on impact. There are no material arrow shafts, metal arrowheads, feathers, or quivers. The bow itself is a solid golden weapon. Stages: 0-3 seconds: In a medium-wide view, Elec is already advancing with a sweeping cane strike. Neith lowers her center of gravity and sidesteps out of its path. While retreating diagonally, she draws the bowstring; a golden-light arrow forms between her drawing hand and the bow. She releases immediately toward Elec. 3-6 seconds: Elec turns his torso to let the first arrow pass beside him and keeps advancing. Neith circles laterally, drawing and releasing a second golden-light arrow without stopping. Elec deflects it with his cane; a small burst of golden sparks marks the point of contact. His follow-through becomes another attack as he closes the distance. 6-9 seconds: At close range, Neith twists away from the cane tip, briefly intercepts the next strike with the solid bow, and steps diagonally outside Elec's reach. Elec turns and follows her movement. Each exchange has a readable approach, contact or evasion, and follow-through. Their hair and garments respond naturally to their movement. 9-12 seconds: Neith springs back to open a short gap, lands with clear foot contact, and continues moving sideways while drawing the bow. Another golden-light arrow forms and launches toward the advancing Elec. The camera moves with them so both the release and the approach remain understandable. 12-15 seconds: Elec brings his cane across the incoming arrow, producing a brief, localized flash at the impact point, and immediately steps into a counterattack. Neith pivots out of his line while bringing the bow into position for her next shot. They are still exchanging attacks as the clip ends, with no victory pose or pause. Visual Style: Detailed cinematic 3D realism, natural skin and fabric, believable material response, and warm directional daylight. Keep fine details clear without excessive sharpening, dense visual noise, or blown-out energy effects. Keep eyes natural, without an added glow. Camera & Performance: One continuous tracking shot. Begin wide enough to establish both fighters and their footing, then make a restrained arc that follows their movement. Ease closer during the close exchange and widen as they separate. Keep the direction of attacks and the space between the fighters legible. Camera changes follow the ongoing action rather than resetting its staging. The stone floor, columns, gateway, and distant cliffs retain their shapes and fixed world positions. Camera movement reveals consistent perspective and parallax. Feet have convincing ground contact; body weight, shadows, hair, and cloth follow the action. No idle aiming, frozen poses, slow motion, teleportation, or sudden background replacement. Audio: Synchronized footsteps on stone, clothing movement, cane-and-bow impacts, a short rising energy sound as each arrow forms, a swift release sound, and a sharp light-energy impact sound. Light environmental wind underneath. No dialogue or narration; keep the physical action clearly audible.

Exclusions: No extra characters, duplicated bodies, costume or equipment swaps, physical arrows, quivers, glowing eyes, contact-sheet layouts, written labels, subtitles, logos, black borders, abrupt scene cuts, warped moving architecture, or effects that hide the fighters.

Plain-text prompt

Actual input images — select an image to inspect the original

PUB-04 · From the first image to the last

1792×1024 · 5.17 seconds · 24 fps · stereo audio

Open original MP4

Measurement / setting	Recorded value
Output	1792×1024 · 124 frames · 24 fps · 5.167 s
Audio	32 kHz stereo
Seed / video shift / audio shift	24681 / 12 / 5
Generation / complete request	551.45 s / 671.34 s
Worker RAM peak (RSS)	30.83 GiB
Owned process-tree RAM peak (RSS / USS / private commit)	32.58 / 26.74 / 39.43 GiB
Generation-worker GPU allocator peak (allocated / reserved; reader excluded)	4.97 / 6.62 GiB
Measured build	rc2-original

The two images below are the actual first/last-frame anchors, with distinct opening and closing actions. Automatic character binding was off.

Prompt — read the exact submitted text

Goal: Connect the supplied opening and closing combat moments with continuous, believable movement over five seconds. Neith keeps her solid golden bow; Elec keeps his cane. Keep exactly two fighters with coherent positions, momentum and equipment. Stages: 0-1.5 seconds: Neith releases an arrow made entirely of golden light. Elec dodges its path and immediately advances instead of waiting for another shot. 1.5-3 seconds: Elec closes the gap while Neith sidesteps and brings the solid bow across her body. Keep their feet grounded and show the change of distance without a cut. 3-5 seconds: Elec steps into a cane strike. Neith braces the solid bow across its path and blocks it, arriving at the closing moment through a readable approach, contact and follow-through. Do not hold a stretched aiming pose. Camera: One continuous shot, tracking gently with the fighters. Preserve the supplied scene, perspective, identities and equipment while revealing natural parallax. Keep both fighters and the point of contact visible; no abrupt reframing. Audio: Synchronized footsteps, clothing movement, an energy release, a brief energy impact and the final cane-and-bow contact, with light wind underneath. No dialogue, narration or music.

Exclusions: No physical arrows, quivers, extra people, glowing eyes, teleportation, equipment swaps, slow motion, frozen poses, scene cuts, warped architecture, contact sheets, labels, subtitles, logos or black borders.

Plain-text prompt

Actual input images — select an image to inspect the original

PUB-05 · Direct 2688×1536 generation — upper-resolution showcase

2688×1536 · 5.17 seconds · 24 fps · stereo audio

Open original MP4

Measurement / setting	Recorded value
Output	2688×1536 · 124 frames · 24 fps · 5.167 s
Audio	32 kHz stereo
Seed / video shift / audio shift	24681 / 12 / 5
Generation / complete request	1,799.97 s / 1,972.19 s
Worker RAM peak (RSS)	33.35 GiB
Owned process-tree RAM peak (RSS / USS / private commit)	36.29 / 30.77 / 43.74 GiB
Generation-worker GPU allocator peak (allocated / reserved; reader excluded)	9.03 / 11.97 GiB
Measured build	rc2-original

An upper-resolution showcase, not the recommended everyday target. I found motion could look strained at this size; 2560×1440 is a suggested setting to try, not a measured alternative here. Physical arrow/quiver elements remain despite the energy-only prompt.

Prompt — read the exact submitted text

Plain-text prompt

Reference images: none — text-only generation.

Verified ComfyUI operation

Current integration: RC.3.5, paired with the unchanged RC.3.1 model. RC.3.5 adds the Windows software installer and retains the in-workflow model preparation introduced in RC.3.4. Its CPU setup checks cover local discovery, full-file verification, pause/resume, failure handling and saved paths in the actual ComfyUI interface. The generation results below retain their tested version identities; this setup update does not introduce a new quality or speed measurement. Setup checks.

The RC.3.1 Safetensors checkpoint and RC.3.3 ComfyUI integration preserve the original source metadata and unchanged learned tensor payloads. The latest RC.3.3 checks used the actual Run button for text-only generation, intentional timeout, cancellation, then successful four-reference recovery with Combat and explicit session release. Both completed MP4 files matched the reference videos I had accepted byte for byte, including audio. Earlier checks retain their original version boundaries. Installed-package checks, recovery and measurements.

The installed APP158b build was exercised directly through the ComfyUI interface on Windows / RTX 4090. The checks included cancellation during generation, a successful request after cancellation, four-image references with Combat LoRA enabled and then disabled in the same worker, visible progress, and worker memory release. The three completed videos passed my quality review. These are separate installation checks, not replacements for the five measured showcase runs above.

A separate 15.08-second, single-speaker sample produced clear English and the complete requested passage, which I confirmed by listening to the result. This demonstrates that sample; it does not promise verbatim speech for every prompt.

Local verification details.

Resource measurements on RTX 4090

All five completed demonstrations are shown here, including the longer and higher-resolution cases. PUB-01: text, 1344×768/5.17 s; PUB-02: four references, 1344×768/5.17 s; PUB-03: four references, 1344×768/15.08 s; PUB-04: first/last images, 1792×1024/5.17 s; PUB-05: text, 2688×1536/5.17 s.

Recorded process memory, in GiB

Sample	Worker RSS	Owned tree RSS	Owned tree USS	Worker GPU allocated	Worker GPU reserved
PUB-01	29.00	29.81	24.31	3.31	3.78
PUB-02	29.72	31.11	25.27	3.81	4.41
PUB-03	32.50	34.00	28.14	7.36	9.72
PUB-04	30.83	32.58	26.74	4.97	6.62
PUB-05	33.35	36.29	30.77	9.03	11.97

RAM covers the recorded Highway workers and owned process tree, excluding unrelated applications. RSS can count shared resident pages more than once; USS counts private resident pages. GPU allocated/reserved are generation-worker PyTorch allocator counters, not total GPU memory. The image reader and allocations outside that allocator are excluded. Reserved includes allocated.

Whole-device VRAM, power, energy and temperature were affected by other applications. I have excluded those readings from the public model-resource tables rather than subtracting an estimated background load. Isolated Highway energy and temperature were not measured in this run set. Complete methodology, private commit and build identities.

Choosing a high-resolution output

For a high-resolution starting point, I suggest 2560×1440. The 2688×1536 example is an upper-resolution showcase and stress case, rather than the recommended everyday target. In my assessment, it can look strained in motion despite completing successfully; more pixels do not necessarily produce a better-looking video.

2560×1440 has not yet been benchmarked in this release-candidate set. This is a practical recommendation to try, not a measured claim that it is sharper, faster or more temporally stable. Use the exact pixel dimensions when choosing settings.

ComfyUI guide · Detailed measurements

Highway V1 builds on MiniMax H3 and the DaSiWa Hybrid v2 INT8 weights. Its contribution is the saved Highway execution architecture and its integrated generation workflow; it is not presented as a newly trained foundation model.

Versioned pair: RC.3.1 Safetensors checkpoint with RC.3.5 ComfyUI integration. Files and checksums · License and attribution · FAQ.

About this project

Highway V1 is an independent project I have developed over roughly a month, building on the work of MiniMax and the wider community. As a solo developer, I cannot test every setup. If something does not work as expected, clear feedback and reproducible reports are especially valuable. Thank you for helping me improve it.

At a glance

	Highway V1
Developer	DucHaiten
Model type	Video-and-audio generation model derived from MiniMax H3
Weight lineage	DaSiWa Hybrid v2 INT8, based on MiniMax H3
Model format	Standard `.safetensors` checkpoint candidate; Highway model definition and transformer weights
User interface	ComfyUI custom nodes and example workflows
Demonstrated dimensions	1344×768 and 1792×1024; 2688×1536 as an upper-resolution showcase
Demonstrated length	124 or 362 frames at 24 fps: approximately 5.17 or 15.08 seconds
Demonstrated audio output	32 kHz stereo, exported with the video
Physical test platform	NVIDIA RTX 4090, Windows; exact environment in the benchmark report

Dimensions and durations above describe the tested combinations below, not every possible cross-product of settings.

What you can create

Text-to-video with audio. Describe subjects, action, camera movement and sound in one request.
Reference-guided scenes. Use multiple images to condition characters, objects and environments. The interface accepts up to nine images; the current demonstration set uses four.
First/last-frame and keyframe generation. Supply visual anchors and describe the transition between them.
Media-conditioned generation and editing. The candidate integration exposes video/audio references, continuation, masks, frame expansion and preservation of source media.
LoRA workflows. Load and remove compatible adapters through the integration. Successful loading does not guarantee the visual quality of every upstream adapter.

The five demonstrations exercise text-to-video, four-image references and first/last-frame generation. The other controls have separate development evidence; they are not all re-evaluated in this demonstration set.

An optional local reference assistant can associate character identities and add grounded clothing cues. The original images still condition generation; they are not replaced by captions. Reference fidelity, clothing retention and object ownership can still fail.

A Safetensors checkpoint with its Highway runtime

The Highway artifact owns the saved model definition, generation policies and its transformer payload. Verified local execution loads that payload without reopening the source H3 transformer checkpoint.

The ComfyUI integration supplies the compatible execution environment and input/output controls. Qwen encoding, video/audio VAEs and media utilities remain explicit dependencies. Consequently, the approximately 20.97 GB transformer artifact is not the size of the complete installation.

Highway V1 retains inherited learned weights. The artifact format and execution changes do not imply new foundation-model training, ownership of upstream weights, or newly learned knowledge. They also do not establish a compression claim by themselves.

The measured high-resolution cases generate at their requested dimensions. They do not first complete a smaller video and then upscale that completed video.

Demonstrations and measured runtime

All five showcase samples passed file and full decode checks and were accepted in my qualitative review. I reviewed this small set myself; it is not a blinded quality study.

Sample	Input mode	Output pixels	Frames / fps	Duration	Request time
PUB-01	Text only	1344×768	124 / 24	5.167 s	359.34 s
PUB-02	Four reference images	1344×768	124 / 24	5.167 s	524.24 s
PUB-03	Four reference images	1344×768	362 / 24	15.083 s	1,473.58 s
PUB-04	First and last images	1792×1024	124 / 24	5.167 s	671.34 s
PUB-05	Text only	2688×1536	124 / 24	5.167 s	1,972.19 s

These are single-run observations on a workstation used for other work at the same time. The reference cases used two distinct reader-fix candidates; the other cases used the original rc.2 artifact. They are not measurements of one finalized release build. Three unsuccessful reader requests are recorded separately.

See the full report for build hashes, generation versus request time, process RAM, worker GPU allocator memory and measurement boundaries. These results do not establish a speedup over another model or minimum hardware requirements.

Using Highway in ComfyUI

The integration includes three portable starting workflows:

Workflow	Purpose
01 — Start Here	A prompt and the core output settings
02 — Image Reference	Upload images and connect them to reference inputs
03 — Complete Controls	Visual controls for conditioning, editing and other optional inputs

There is no separate Highway desktop app. A compatible Highway integration is required; the integration loads the Highway architecture and its saved generation policy.

The ComfyUI guide documents installation, dependencies and controls for the versioned pair. Use the supplied versioned package; this guide does not rely on a ComfyUI Manager entry.

Community development

Highway's community interface documents 368 logical LoRA targets and 52 packed QKV aliases, with exact names, shapes and scaling rules. A small CPU-only preflight tool helps authors check adapter files without loading the model. Workflows, conversion tools and independent training research can build around documented interfaces under the applicable terms.

The current runtime supports compatible LoRA inference. A complete gradient-enabled trainer and full fine-tuned-model export path are not supplied or qualified by this release candidate. Community status and interface kit · Technical overview.

Intended use and limitations

Highway is intended for creative video production, prototyping and research into generative media, subject to the applicable component licenses and rights in the inputs.

Characters, clothes, equipment, motion, physical interactions and scene continuity can be incorrect. Prompt exclusions are not guarantees: physical arrows and a quiver appeared in the text-only demonstrations despite requests for light-only projectiles.
Four-image conditioning does not prove perfect semantic assignment of every image or exclusive ownership of every object. In the measured reference cases, two character associations were confirmed; the environment and equipment associations remained unresolved.
No broad language, dialogue, social-bias or safety evaluation is reported here. Synthetic scenes should not be treated as evidence of real events.
The measured environment is Windows with an RTX 4090. Results and requirements on this page refer to that tested configuration.
No universal minimum RAM/VRAM, cross-device speed guarantee, or guaranteed energy/temperature reduction is established. Changing the output frame rate is not evidence of native 60 fps motion capability.

Provenance and licensing

The immediate transformer source is DaSiWa Hybrid v2 INT8, model 2877206 / version 3314675, based on MiniMax H3. I develop Highway under the name DucHaiten. This is not an official MiniMax release or an endorsement by the upstream authors.

The MiniMax H3 license and Acceptable Use Policy, source notices and component-specific terms apply. Keep the accompanying NOTICE and attribution when redistributing. Territory, recipient-use and commercial conditions are summarized in License and attribution.

I would like to thank the MiniMax H3 and DaSiWa authors, WanGP, Qwen, ComfyUI, and the numerical and media libraries used by the implementation. Each component retains its own rights and terms; this documentation does not relicense them.

License and attribution · Frequently asked questions. Component licenses remain distinct; no unrestricted open-source or worldwide commercial grant is implied.

ComfyUI update: RC.3.6

Unused standard image-upload slots are now skipped automatically; keep them connected for later use. Update the ComfyUI integration, rerun its installer with ComfyUI closed, then restart. Existing workflows and model files can be reused.

RC.3.8: model discovery and support diagnostics

Reuse models from shared/linked folders and custom locations, including renamed identical weights. Download a support report from the setup panel when something fails. Update and troubleshooting details. Checkpoint and generation quality are unchanged.

Downloads last month: -; Downloads are not tracked for this model. How to track