✨ You asked, we delivered! To everyone running on an 8GB VRAM setup who thought they were locked out of top-tier video gen: this one is for you. 🤝
This advanced, two-stage workflow completely solves the budget VRAM crisis. It harnesses Minimax’s incredible motion adherence and base generation power as your foundation, then instantly hands off the baton to LTXV acting as a high-fidelity refiner.
The result? Crisp, high-resolution cinematic videos without triggering an Out-Of-Memory (OOM) error! 🤯
While this combo workflow has unmatched quality and fidelity, we've released a purely Minimax workflow. Check it out here. - https://civarchive.com/models/2941580/minimax-solo-int8-workflow-pure-8gb-power-low-storage-low-vram
V4 is in works. We're going to make this even better. READ THE CHANGELOG for list of Updates.
🛠️ Changelog & Bug Fixes
Known problem in V1.0: First few frames ignored by LTX Refiner , Artifacts produced in the Minimax stage cause Turbo LoRA's at the moment are not perfect.
Fixed in V2.0: Now your final output will have clarity and quality from the first frame itself. Also, LTX Refiner will fix the anatomy issues as well as any artifacts along with refining the video. Also, you can now use Interpolation along with RTX without much delay with Clean RAM VRAM nodes being added to the groups. A note is there in the workflow on how to connect RIFE Interpolation with RTX Upscaling.
Fixed in V3.0: Lower resolutions not refining well now give crystal clear output. Clarity has increased atleast 2-3x more. Added 2 loops for video extension. Unlimited loop video possibility. Switches given for streamlined rapid workflow as per requirements. Upscale Model stage added. T2V Mode is available now. Added another LoRA for LTX to refine videos. Get it from here - https://huggingface.co/CQdesign/LTX-2.5-CQ-Video-and-Image-Enhancer-LoRAs/blob/main/ltx2.5-CQ-enhancer-lora-for-videos-rank128.safetensors
🎨 The Feature Stack
8GB VRAM Friendly: Optimized explicitly for low-budget hardware using extreme INT4/INT8 quantization.
Two-Stage Refinement: Uses Minimax for raw motion framework, and LTXV to iron out details.
Smooth Frame Interpolation: 2x FPS motion smoothing baked right into the output line.
4K Upscaling Power: Blast your final render all the way up to sharp 4K resolution.
Backward Compatible: Don't have LTX 2.5? This pipeline is mostly backward-compatible with LTX 2.3 components!
🧠 The Shopping List: Required Models
To prevent errors, make sure you have these exact files placed in your ComfyUI directories:
⚡ Stage 1: The Minimax Engine
Checkpoint:
MinimaxH3INT8INT4_fl2vaINT8Pruned.safetensorsText Encoder:
Qwen3vl_32b_minimax_h3_int4_convrot.safetensorsVAEs:
minimax_h3_video_vae_fp16.safetensors&minimax_h3_audio_vae_fp32.safetensorsLoRA:
minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors
🌀 Stage 2: The LTXV Refiner & Upscaler (LTX 2.5 or 2.3)
VAEs:
LTX-2.5-audio-vae-bf16.safetensors<X-2.5-video-vae-conv-bf16.safetensorsTransformer:
LTX-2.5-22b-distilled-transformer-comfy-in8-convrot.safetensorsText Encoders:
Gemma4-12b-with-proj-ltx-2.5-comfyui-int8-convrot.safetensorsSpatial Upscaler:
LTX-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensorsLoRA:
LTX2.3_Crisp_Enhance.safetensors & ltx2.5-CQ-enhancer-lora-for-videos-rank128.safetensors
🛠️ Quick Start
Import the workflow file into your workspace.
Verify all paths point correctly to your quantized INT4/INT8 checkpoints.
Queue Prompt and enjoy elite video generation right on your local 8GB card!
💬 Join the Low VRAM Movement!
If this setup saves your GPU from bursting into flames, hit that Like button and leave a review!
👇 Low VRAM doesn't mean low quality! Drop your 4K interpolated video renders and test clips in the comments below! 👇
Description
FAQ
Comments (11)
can this be used as t2v only?
Originally, no. But now you can, if you download the updated workflow. Needed slight tweaks and addition of a switch. There are notes in the workflow to get you started with T2V.
OMG, my GPU isn’t screaming from overheating anymore! Thank you so much for all the effort. Is this workflow also compatible with 16:9 landscape video?
All 9:16 resolutions will work when reversed for 16:9
You're welcome. Cheers. We won't let them keep us out. Consider showing the support by liking the workflow, resource and sharing your outputs here. Can't wait to see the older/Low VRAM GPUs shining through this.
I'm not paywalling my work, that shouldn't mean people are going to be ungrateful & can't even bother a like. CivitAI is heavily biased towards creators with thousands of followers & those followers have herd mentality to treat those creators like celebs creating situations where those creators are paywalling their models LoRAs or workflows with close to million buzz. You're feeding the wrong creators. I've literally released my 1 year worth of work to create refined and heavily tested workflows for free & continue to release more versions. There are no beta versions here to make you frustrated for not being able to run them & the results are guaranteed. I've also left prompts in there that if you understand would change your video outputs drastically since its all about direction from text conditioning with most of the models.
我支持你，bro
能告诉我 你呈现的视频的生成时间花废了多久，我的笔记本，的硬盘不够了。之前我试过15秒的视频，864x480，6step，需要，600秒的工作流。这个工作流的人物皮肤很舒服，如果我要按照你的工作流，需要多少的硬盘空间，生成的时间有是多少
我不是懒惰，我的硬盘不够空间，我需要甄选。我之前的视频都是，要速度没质量，要质量就要很久（其实质量还是一般）
@mayih8511823 Hi this is a rapid workflow with 4 step Turbo Lora. Since We're using both Minimax & LTX 2.5, we will need both their model clip vae LoRA together. The models clips used are INT8/INT4 precision hence require less space. In total 82GB disk space would be needed to accommodate this setup. I recommend always ensuring 100GB at least as empty disc space & putting that as paging file to keep your system free of any lag while working on any high workload.
Now for your reference
Workflow in latest SonicLoopsV3
Minimax H3 0.5 -> LTX 2.5 REFINE -> 2X UPSCALE MODEL -> 1.5X RTX UPSCALE ( Interpolation option between the two upscale stages bypassed )
832 x 448 - 15 Seconds - 24 FPS - 360 Frames - Final Resolution 2496 x 1344 - 398 Seconds
For the above video
1024 X 768 - 7 seconds - 24FPS - 168 Frames - Final Resolution 3072 x 1728 - 313 Seconds
You may choose to use 6 step with the Turbo LoRA you want to use, it will barely add 20-30 seconds more to this in both cases. Rest assured you will get both quality and speed. Use only the resolutions given in the workflow. Test and share your outputs as well as your generation time. It will be good feedback for V4.
首先在这里非常感谢作者的分享，通过我本地的agent修改了一些参数，终于是跑出了，5秒钟的视频 花费了6分钟，但是得出的质量令我挺满意的，他对动漫的清晰度非常高，很清晰的还原了我的参考图片，但是在第一个生成的视频我看到了模糊 还有闪光，这导致了我后面的修复 还是有烟花一样的，就把它当做特效吧😂，或许是我的提示词，我只是简单的描述，让眼睛动起来
It has to do with the Turbo LoRA you might be using for 6 step, at this point the default LoRA included with the workflow is the most stable and must be used with 4 step only. There is really no need of tweaking as the best settings have been applied which will work across realism and anime both. To make the workflow even faster you may choose to opt not running the upscale model as well and directly use RTX to upscale at 2x/4x, that would reduce 1-2 minutes further for production. This is a low VRAM setup for those who can't run Minimax directly at the higher resolutions.