Instructions to use lightx2v/Minimax-h3-Turbo-SLA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Inference
- Notebooks
- Google Colab
- Kaggle
Testing turbo LoRAs
Testing of 3 Turbo LoRA against official H3 and LightX2V documentation
- Official H3 FL2V workflow with SageAttention and Shift nodes added
- Official models used
- Official prompt guide used
- Settings: 1 MP, 15 seconds, 4 steps, euler, simple, 6video 3audio shifts, fixed seed. Used RTX 5090. All gens basically 44s/it
=================
Video 0 - baseline: No Turbo LoRAs - 50 steps
=================
Video 1: v0.1 LightX2V Turbo SLA LoRA
=================
Video 2: Silveroxides dareties Turbo LoRA
=================
Video 3: v1.1 LightX2V Turbo 768 LoRA
=================
Video 4: old LightX2V 4-step v0.1 LoRA
I decided to include this since people liked it a lot. I see why; its prompt adherence is the best from any turbo LoRA; followed everything correctly and physics looks correct everything - though quality was worsened.
=================
EXTRA:
Video 5: Silveroxides dareties Turbo LoRA with Hybrid B30 model instead of official H3 int8 FL2V model
=================
EXTRA no2:
Video 6: v1.1 LightX2V Turbo 768 LoRA with video_shift 8 instead
The prompt adherence became much better with video_shift 8 - followed everything in the prompt: no lip talking in the beginning, hand has 5 fingers and thumb over the spider as prompted.
This might be the best of all.
Notes:
Baseline is obviously the best, the only video that does not make the girl talk with her lips, as the prompt says: "She (S1) says in an off-screen voiceover: ...".
But it failed at adhering to prompt: "... her thumb poised just above the spider. The camera holds a static shot as her thumb hovers and tightens, then slowly relaxes and draws back without crushing it", she used other index finger.
Prompt was:
How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 5) aligns with the 15.08-second mark of the target video.
integrated_multimodal_description: [Shot 1] Live-action, cinematic, ultra-realistic, a cool softly-lit palette of pale daylight from a viewport against warm amber accent strips, with shallow depth of field. The video begins on the exact framing established by Picture 1: a front-facing close-up of the platinum-blonde woman inside the dim interior of a futuristic ship, her long wavy white hair loosely braided at the temples, green eyes level with the lens, light freckles across her nose and cheeks, a gold filigree forehead headpiece set with an iridescent multicolour gemstone, matching gold drop earrings, a white lace high collar, gold filigree shoulder pieces, and a pearlescent iridescent bodice with gold scrollwork and coloured gems, framed by dark angular hull panels, amber light strips overhead, and a wide viewport of pale clouded sky to her left. The camera pulls out with small amplitude at slow speed as she turns her head a few degrees and lets her gaze drift toward the viewport, cold light sliding across the gemstones on her bodice. The woman, in a low, controlled voice speaking English with a pronounced French accent (S1), says in an off-screen voiceover: <d>[English] When I was a girl... I had a fear of spiders...</d> while her lips remain completely closed. [Shot 2] At 00:02.800, the shot cuts to an extreme close-up of her open palm holding a small dark spider, the gold filigree of her shoulder piece soft and out of focus at the edge of frame, her thumb poised just above the spider. The camera holds a static shot as her thumb hovers and tightens, then slowly relaxes and draws back without crushing it; the spider stays on her palm and shifts its legs. She (S1) says in an off-screen voiceover: <d>[English] I was told that they felt no emotion...</d> while her lips remain completely closed out of frame. [Shot 3] At 00:06.000, the shot cuts to an exterior wide of the ship's dark hull banking away high above a city under a pale overcast sky, thin cloud tearing past the wings. The camera arcs with large amplitude at slow speed around the banking hull toward an open bay, finding her silhouette standing in the bay doorway as the heavy sliding doors close toward each other, wind dragging at her hair and the hem of her collar while she stares back at the receding city. She (S1) says in an off-screen voiceover: <d>[English] ...that their hearts never beat. But I know the truth...</d> while her lips remain completely closed. [Shot 4] At 00:09.200, the shot cuts to an extreme close-up of her single green eye, freckled skin, pale lashes, and the lower edge of the gold headpiece filling the frame. She (S1) says in an off-screen voiceover: <d>[English] At the moment of the kill...</d> while her lips remain completely closed. At 00:10.800 she blinks once, and as the lid opens the green iris flashes with a brief red glitch, a fractured red data-streak tearing across the pupil for a fraction of a second before the iris returns to green, landing exactly on the percussive stinger in the score. [Shot 5] At 00:12.200, the shot cuts back to the front-facing close-up matching the head angle, hair fall, and hull framing established by Picture 1. The camera pushes in with small amplitude at slow speed as she settles perfectly still and her expression hardens, the differences from the final framing progressively narrowing. She (S1) speaks directly to camera with her lips moving: <d>[English] They are never more alive.</d> On the final word both eyes ignite with a slow red glow that rises and pulses once, and the framing, head angle, lighting, and glowing red eyes settle into the exact composition established by Picture 2, holding to the end of the video.
overall_soundscape: A low hull hum and ventilation drone fill the ship interior, punctuated by crisp mechanical clicks and small console relays. Fabric shifts faintly as her thumb moves above her palm. Wind rushes past the open bay before the heavy doors grind and boom shut, and her breathing stays slow and shallow throughout.
non_diegetic_music: A sustained low synthesizer pad and bass drone at a slow tempo, increasing steadily in volume across the video. A short percussive stinger hits at the blink near 00:10.800, and heavy low drum strikes at a slow heartbeat rhythm enter in the second half. The strikes space further apart and resolve on a single downbeat under the final line.
Maybe it's not working correctly in comfy, because it's actually slower for me
Maybe it's not working correctly in comfy, because it's actually slower for me
They're all the same speed for me (44s/it).
ComfyUI doesn’t have an SLA backend, right? And as far as I understand, the speed would be about the same as the regular version, correct?
ComfyUI doesn’t have an SLA backend, right? And as far as I understand, the speed would be about the same as the regular version, correct?
ComfyUI with no SLA backend is right, hence why no speedups for us comfy users.
Great, which one is the fastest?
Thanks
my nodepack contains the node for enabling sparse attention, the lora is not required.
really thanks! look forward the r2va!

