Online Workflow Experience: MiniMax H3 – Cinematic Version (Official Release)
Try it here: https://www.runninghub.ai/zh-cn/post/2092446042583642113/?inviteCode=rh-v1635
MiniMax H3 Cinematic LoRA (Official Version)
Recommended Model Strength: 0.7 for standard shots; around 0.5 for high-motion sequences. For noise-specific inputs: 0.6 for low noise, 0.5 for high noise. The trigger word has been removed; no need to input "DY".
The MiniMax H3 Cinematic LoRA (Official Version) builds upon the original model's authentic cinematic texture with further optimizations. It goes beyond simply enhancing the "cinematic feel" of the image, focusing instead on specific improvements to address common issues encountered during actual video generation.
This version further enhances the lighting depth, color relationships, skin texture, depth of field, and atmospheric quality characteristic of real cinematography. It places special emphasis on stabilizing character limbs and ensuring natural movement, thereby reducing issues such as limb distortion and awkward motion. Targeted improvements have also been made to address H3's tendency to trigger unwanted automatic cuts, sudden camera jumps, or breaks in continuity during long takes, resulting in smoother continuous camera movement and more cohesive long shots.
Additionally, optimizations have been made to mitigate motion artifacts—such as streaking, ghosting, and image breakdown—in high-motion scenes involving running, fighting, chases, or rapid camera movement. This ensures better character definition and image stability during high-speed action while preserving the intended range of motion and sense of speed.
While retaining MiniMax H3’s original capabilities regarding motion, camera understanding, character performance, and audio-visual generation, the output is now more natural, restrained, and realistic. It minimizes "plastic" textures, over-sharpening, harsh lighting, and obvious AI artifacts, resulting in visuals that closely resemble actual cinematic footage rather than a simple overlay of a film filter.
It is suitable for a wide range of scenarios, including continuous long takes, dialogue-heavy scenes, nuanced emotional performances, youth-oriented films, narrative shorts, chase and fight sequences, high-motion shots, and genres such as Wuxia, urban drama, and suspense, as well as cinematic advertisements. It also serves as a versatile LoRA for enhancing the cinematic quality of everyday MiniMax H3 generations.
Core Upgrades in the Official Version: More stable character limbs | Enhanced cinematic quality | Improved handling of automatic cuts | Stronger continuity for long takes | Optimized handling of high-motion streaking and ghosting | No trigger word required—ready to use immediately.
Description
Online Workflow Experience: MiniMax H3 – Cinematic Version (Official Release)
Try it here: https://www.runninghub.ai/zh-cn/post/2092446042583642113/?inviteCode=rh-v1635
MiniMax H3 Cinematic LoRA (Official Version)
Recommended Model Strength: 0.7 for standard shots; around 0.5 for high-motion sequences. For noise-specific inputs: 0.6 for low noise, 0.5 for high noise. The trigger word has been removed; no need to input "DY".
The MiniMax H3 Cinematic LoRA (Official Version) builds upon the original model's authentic cinematic texture with further optimizations. It goes beyond simply enhancing the "cinematic feel" of the image, focusing instead on specific improvements to address common issues encountered during actual video generation.
This version further enhances the lighting depth, color relationships, skin texture, depth of field, and atmospheric quality characteristic of real cinematography. It places special emphasis on stabilizing character limbs and ensuring natural movement, thereby reducing issues such as limb distortion and awkward motion. Targeted improvements have also been made to address H3's tendency to trigger unwanted automatic cuts, sudden camera jumps, or breaks in continuity during long takes, resulting in smoother continuous camera movement and more cohesive long shots.
Additionally, optimizations have been made to mitigate motion artifacts—such as streaking, ghosting, and image breakdown—in high-motion scenes involving running, fighting, chases, or rapid camera movement. This ensures better character definition and image stability during high-speed action while preserving the intended range of motion and sense of speed.
While retaining MiniMax H3’s original capabilities regarding motion, camera understanding, character performance, and audio-visual generation, the output is now more natural, restrained, and realistic. It minimizes "plastic" textures, over-sharpening, harsh lighting, and obvious AI artifacts, resulting in visuals that closely resemble actual cinematic footage rather than a simple overlay of a film filter.
It is suitable for a wide range of scenarios, including continuous long takes, dialogue-heavy scenes, nuanced emotional performances, youth-oriented films, narrative shorts, chase and fight sequences, high-motion shots, and genres such as Wuxia, urban drama, and suspense, as well as cinematic advertisements. It also serves as a versatile LoRA for enhancing the cinematic quality of everyday MiniMax H3 generations.
Core Upgrades in the Official Version: More stable character limbs | Enhanced cinematic quality | Improved handling of automatic cuts | Stronger continuity for long takes | Optimized handling of high-motion streaking and ghosting | No trigger word required—ready to use immediately.
FAQ
Comments (10)
woooo,感谢你分享它
So you released a version with no tensor errors, then released a version of the same lora WITH tensor errors? Why?
Good question.
yeap what is the point!?
Compared to the version that suppresses tensor-related error messages, this one uses the original model parameters without any trimming; although it may trigger error messages, this does not affect functionality. I am simply offering you more options—okay?
@TuTu_1018 so the question is, which one is better?
@mankogamerhn The original version, of course.
I've been asking the same question like why don't just use a single post and then upload several version lmao, btw i've been testing the lora and i prefer this https://civitai.red/models/2849726/cinematic-style-lora-detail-enhancer-for-minimax-h3?modelVersionId=3217898 cause it really understand cinematography and camera term.
@LeakNudeCollector You like version 1 specifically? I did a test of both. They do each have different strengths. Version 2 had definite improvements when not using turbo loras, but it was actually were worse than no lora if using turbo. I didn't do a big test of version 1, though I did include it in comparisons.
The LoRA is okay but lots of static shots where the camera should be moving, like drive-by-shots, no LoRA looks better for those. I suspect, a lot of steadicam in the dataset.