Cinematic Style V2 for MiniMax H3
Cinema V2 is a major expansion of my broad cinematic style LoRA for MiniMax H3.
The dataset now contains 1,852 cinematic images with 1,852 matching detailed captions — 3,704 files in total. Version 1 already provided a strong general foundation for realistic film photography, framing, lighting, lens character, depth of field, production design, atmosphere, realistic materials and organic film texture. Version 2 keeps that foundation and adds much more deliberate camera-position and composition training.
## What Is New in V2
The dataset was expanded from 1,201 to 1,852 image-caption pairs, adding 651 new cinematic examples.
New dedicated categories include:
- Aerial shots
- Clean single-subject compositions
- Dutch angles
- Establishing shots
- High-angle shots
- Low-angle shots
- Over-the-shoulder compositions
- Overhead shots
These additions improve the LoRA's understanding of camera height, viewing direction, spatial depth, visual geometry, subject placement, negative space and environmental scale.
The original dataset already covered the main shot sizes:
- Extreme Close-Up
- Close-Up
- Medium Close-Up
- Medium
- Medium Wide
- Wide
- Extreme Wide
Together, both parts now provide a much broader cinematic vocabulary. V2 is better suited for quiet feature-film scenes, atmospheric openers, dialogue scenes, character moments, environmental storytelling, news-style cinematography, trailers, science fiction, period drama, crime, romance, horror and large-scale landscapes.
The captions place strong emphasis on:
- exact framing and subject position
- motivated key, fill, rim and practical lighting
- lens character and depth of field
- realistic skin, fabric, glass, metal and wet-surface response
- foreground layering and environmental depth
- controlled color relationships
- film grain, halation, bloom and highlight rolloff
- atmosphere, weather and production design
This is a cinematography style LoRA, not a character LoRA and not one fixed movie look. The genre, people, clothing, environment and action remain controlled by your prompt.
## Activation Tag
Use the original activation tag exactly as trained:
ASTROCINEMAV01K2T
The K2T wording is intentional and must not be changed. This release is trained for MiniMax H3.
Place the trigger near the beginning of integrated_multimodal_description.
Example:
```text
integrated_multimodal_description: [Shot 1] ASTROCINEMAV01K2T depicts a live-action cinematic scene in a native 16:9 widescreen frame, photographed on a 50 mm anamorphic lens with motivated tungsten practical lighting, cool window fill, realistic skin texture, shallow depth of field, fine 35 mm grain, subtle halation and controlled highlight rolloff. No cuts; one continuous shot throughout. Describe the complete subject, wardrobe, environment, blocking, camera movement and spoken dialogue here.
overall_soundscape: Describe the synchronized ambience, physical sounds and nonverbal character sounds here.
non_diegetic_music: Describe the score and instrumentation here, or use N/A when the scene should rely only on natural sound.
```
## Important
The LoRA strengthens the visual cinematic language: framing, camera placement, perspective, lighting, lens feeling, production design and film texture. MiniMax H3 itself remains responsible for temporal motion, acting, speech, sound effects and music, so those elements should still be described clearly in the prompt.
For the best results, do not write only “cinematic.” Direct the scene like a filmmaker: specify the shot size, camera height, angle, lens, movement, light sources, subject blocking, environment, atmosphere and sound.
_______________________________________________________________________
# Cinematic Style LoRA for MiniMax H3
A broad cinematic style LoRA for MiniMax H3, trained on roughly 1,200 carefully captioned cinematic stills.
The goal is not to force one specific movie look. The dataset was built to teach a wider cinematic language: realistic film photography, composition, framing, perspective, motivated lighting, lens character, depth of field, production design, realistic materials and skin, color relationships, atmosphere and organic film texture.
## Activation Tag
Use the original activation tag exactly as trained:
ASTROCINEMAV01K2TYes, the tag still contains K2T. That is intentional because it is the original learned activation tag from the dataset. This release itself is for MiniMax H3.
Place the tag near the beginning of style_and_tone.
Example:
style_and_tone: ASTROCINEMAV01K2T. Photorealistic neo-noir crime drama, motivated practical lighting, shallow depth of field, restrained color grading, realistic skin texture, subtle halation and organic fine film grain.## How I Recommend Prompting It in MiniMax H3
Do not rely only on the word "cinematic".
Tell H3 what kind of cinematography you actually want:
- time of day
- dominant and secondary light sources
- practical lights visible in the scene
- camera distance and framing
- camera movement
- lens feeling / depth of field
- realistic skin and material response
- environment and production design
- weather and atmosphere
- color relationship
- film grain, halation or highlight rolloff when useful
The LoRA works best as a broad cinematic foundation, while the rest of the prompt defines the specific genre and shot.
## MiniMax H3 Prompt Structure
I recommend separating the prompt into:
- style_and_tone = visual language and cinematography
- continuous_shot or multi_shot_sequence = action, acting, camera and timing
- audio_design = music, sound effects and ambience
- explicit dialogue lines whenever characters speak
H3 is very good at combining visual direction with audio, so do not forget the audio side when the scene needs it.
For example, a crime scene can include rain, passing traffic, a low score and dialogue. A fantasy battle can include orchestral music, weapon impacts, footsteps and creature sounds. An advertisement can use voice-over, product Foley and a clean commercial music bed.
## General Tips
For realistic cinematic results, avoid contradictory style language. If you want realism, do not mix the prompt with vector-art, cel-shading or illustration terminology.
Use physically coherent lighting. Instead of listing random light colors, describe which source is dominant and how secondary light affects the subject.
Keep skin and materials realistic. Describe pores, tonal variation, fabric response, wet surfaces, polished metal, glass or practical reflections only when they are relevant to the shot.
For fast action, keep the movement readable. One clear action per beat usually works better than trying to force too many unrelated events into a few seconds.
For trailers and advertisements, multi_shot_sequence is extremely useful because each cut can have a clear purpose.
For music-led videos, describe the music as the master timeline. Shot and camera cuts may use timecodes, but avoid forcing sung lyric lines into strict timed segments.
## What to Test
The LoRA is useful for a wide range of H3 projects:
- crime and neo-noir
- period drama
- romance
- action
- science fiction
- fantasy RPG cinematics
- realistic game trailers
- commercials
- beauty advertising
- product videos
- music videos
- atmospheric character scenes
- cinematic environments
The attached example videos were generated with MiniMax H3 using this LoRA. I also included a separate prompt pack containing 15 complete H3 Text-to-Video prompts across film, games, romance, action, trailers, advertising and music-video scenarios.
Feel free to modify the examples and push the LoRA into completely different genres. If you create something interesting, I would love to see the result.
Description
Cinematic Style V2 for MiniMax H3
Cinema V2 is a major expansion of my broad cinematic style LoRA for MiniMax H3.
The dataset now contains 1,852 cinematic images with 1,852 matching detailed captions — 3,704 files in total. Version 1 already provided a strong general foundation for realistic film photography, framing, lighting, lens character, depth of field, production design, atmosphere, realistic materials and organic film texture. Version 2 keeps that foundation and adds much more deliberate camera-position and composition training.
## What Is New in V2
The dataset was expanded from 1,201 to 1,852 image-caption pairs, adding 651 new cinematic examples.
New dedicated categories include:
- Aerial shots
- Clean single-subject compositions
- Dutch angles
- Establishing shots
- High-angle shots
- Low-angle shots
- Over-the-shoulder compositions
- Overhead shots
These additions improve the LoRA's understanding of camera height, viewing direction, spatial depth, visual geometry, subject placement, negative space and environmental scale.
The original dataset already covered the main shot sizes:
- Extreme Close-Up
- Close-Up
- Medium Close-Up
- Medium
- Medium Wide
- Wide
- Extreme Wide
Together, both parts now provide a much broader cinematic vocabulary. V2 is better suited for quiet feature-film scenes, atmospheric openers, dialogue scenes, character moments, environmental storytelling, news-style cinematography, trailers, science fiction, period drama, crime, romance, horror and large-scale landscapes.
The captions place strong emphasis on:
- exact framing and subject position
- motivated key, fill, rim and practical lighting
- lens character and depth of field
- realistic skin, fabric, glass, metal and wet-surface response
- foreground layering and environmental depth
- controlled color relationships
- film grain, halation, bloom and highlight rolloff
- atmosphere, weather and production design
This is a cinematography style LoRA, not a character LoRA and not one fixed movie look. The genre, people, clothing, environment and action remain controlled by your prompt.
## Activation Tag
Use the original activation tag exactly as trained:
ASTROCINEMAV01K2T
The K2T wording is intentional and must not be changed. This release is trained for MiniMax H3.
Place the trigger near the beginning of integrated_multimodal_description.
Example:
```text
integrated_multimodal_description: [Shot 1] ASTROCINEMAV01K2T depicts a live-action cinematic scene in a native 16:9 widescreen frame, photographed on a 50 mm anamorphic lens with motivated tungsten practical lighting, cool window fill, realistic skin texture, shallow depth of field, fine 35 mm grain, subtle halation and controlled highlight rolloff. No cuts; one continuous shot throughout. Describe the complete subject, wardrobe, environment, blocking, camera movement and spoken dialogue here.
overall_soundscape: Describe the synchronized ambience, physical sounds and nonverbal character sounds here.
non_diegetic_music: Describe the score and instrumentation here, or use N/A when the scene should rely only on natural sound.
```
## Important
The LoRA strengthens the visual cinematic language: framing, camera placement, perspective, lighting, lens feeling, production design and film texture. MiniMax H3 itself remains responsible for temporal motion, acting, speech, sound effects and music, so those elements should still be described clearly in the prompt.
For the best results, do not write only “cinematic.” Direct the scene like a filmmaker: specify the shot size, camera height, angle, lens, movement, light sources, subject blocking, environment, atmosphere and sound.
FAQ
Comments (60)
I have a question, do I have to use both versions of the loras at the same time? Or both loras are merged together on the updated version? I really love this lora! Thank you for your hard work!
no V2.0 is everything from V01 + Detail enhancer and extended knowledge in camera angles and format
@Astroburner great! I'll try to buy yellow buzz, but I don't know how to do it. I made a transfer, but still haven't received the buzz. I hope I did everything good. Thank you very much for your reply. I will buy if someday I receive my buzz.
@mankogamerhn i give you 1000 as present, xD i send to you
@Astroburner thank you very much! You're amazing just like your lora! I have no words to explain how much. I already purchased your lora. I will try it now!
OK this version is looking very good in the examples cant wait to try it
I really want to try it out, but there is absolutely no way to purchase the Yellow Buzz. Is there any way to do this?
sure take my 500 buzz xD
@Astroburner
I can't believe how grateful I am!! Thank you from the bottom of my heart!
@kennedysworks No problem, have fun with it. If you need help just write me xD
Thanks for the tips! This LoRA is great, highly recommend. Double thumbs up! :D
Edit: I made comparison videos, and posted them, using the example prompts, to ensure that it pulls out what was trained into the loras, and it sure seems to do that. I included version 1 in my comparisons as well, for the curious. Might be a big ask, but DON'T use turbo loras. You won't be getting what it was trained to do.
------
Now that they're charging for the lora, could someone do a comparison video of with and without? I've gotten some clearer video that looks like film out of stock H3, but I'd love to see what exactly this adds. The examples look a bit blurry compared to what I've made without it or any loras (except turbo), so I want to see if it's worth buying. Yellow buzz is hard to get. Thank you!
Could be wrong but the blur and noisiness is apart of the cinematic style. Crisp clean generally isn't "cinematic".
@pumaai487 Yeah, films are filled with some of the best texture out there. They do have a focal point, and some parts may be blurry due to being out of focus, but that comes from choosing where to focus, and that part as very clear. If you are watching a film and it's out of focus, you got a bad transfer/upscale. You might be watching it on VHS too.
The quality I'm seeing in the main examples looks like a phone from the 2010s. That might be fine, but I've gotten much cleaner and sharper videos out of Minimax without any loras except for turbo lora, which should lower the quality, not raise it. Check out my videos. They're not HD, but they're clearer, and have better texture. I don't want to judge this lora on those example videos.
@Jellai the LoRA understands better frames, shot types and camera angles. Also different color gradings and filters. As well better 16:9 screen play.
@Astroburner ah. I guess that's why I wanted to see comparisons, so that I understand how much better it understands those things. Normally I'd just download and test it myself, but I have the yellow buzz hurdle this time, so I have to find out if it's worth doing the leg work to get some.
But thinking about it now, maybe I could run tests without the lora and just compare with your videos. I'll put together a comparison. I'm just afraid mine will look better due to my setup, since your examples are blurry. I don't want to accidentally make the Lora look bad. You didn't use starting images or reference, right?
@Jellai I send you 500 Buzz feel free to test it now ;-)
@Astroburner thank you! I'll do some tests today.
@Astroburner So far, I'm finding that it really shines with euler/simple. I was able to get some good/better results out of H3 with other schedulers/samplers, but with this lora specifically, the cinematic lora fell short when using non euler/simple, but certainly shined over the base model when using that. It's also possible that it suffers under turbo more than the base model. But yeah, using non-turbo, euler/simple, it was a clear improvement. I have to run more of those tests in particular though, under some different scenes.
@Jellai Had some Bad results on a turbo , Like 30/70 Ratio. Without turbo in res_multi only great results, but longer waiting time xD
@Astroburner @Jellai I think most people may think of turbo making things have less detail, and maybe effecting prompt adherence in subtle ways. But turbo lora can really kill your camera blocking and framing instructions, and seems to prevent getting anything close to an Extreme Close-Up. Like it won't even do it, and backs the camera off, framing wider, sometimes back to a medium close-up with the same prompt, same seed. Then same seed, turn off turbo, and now you get a frame composed more how you were directing it, and more pointing the direction you might have asked for.
Sampler/Scheduler does also effect these things, much more than I had previously thought. Anymore I seem to get weird anatomy and morph/dissolve in the background when I use Euler, though I'd had okay luck with Euler/lin_quad before, for unchallenging wides and mediums. Or er_sde/power_shift. But they tend to inhibit other things, while looking clean.
For ref2v at least, I'm back to res_multi/simp for the time being, much as I don't want to, but it is the closest to giving me an actual ECU and following some camera direction using a ref environment.
@Astroburner I posted my comparison videos. Used your prompts to make sure it's tapping into what you trained. Unfortunately, the alien news video, using your alien news prompt, got flagged for review, rated X. So weird.
@pumaai487 they might be too young to know the difference.
Only way to make a modern digital cinema camera not look like expensive camcorder/smartphone footage is to throw some classic film era lenses on it. Which will tend to be softer, more dimensional and have character and texture to them.
@BRai4now I'm not young, and I know the difference. I'm so old that I'm thinking about film, rather than high resolution digital cameras. The amount of detail that can be pulled out of film is astonishing. We've been trying to catch up to what we abandoned for decades.
@Jellai cinema lenses tend to be softer. They're shooting to a smaller size 35mm, run vertically, not your SLR film, and with bigger aperture f-stop than 35mm photography outside of the "street photography" context. And in the final days of film-based cinema it was the norm most movies were shot "Super 35", and seeing these released on 4K is pretty dubious, because that stuff wasn't worth scanning 4K for post production, working at higher fidelity than you've ever seen 35mm motion picture film in your life.
What theoretically "could" be achieved with perfect studio conditions and small aperture f-stop and really boring lenses isn't relevant in the context of this lora or "cinematic" in general.
Shooting cinema lenses with stops between t1.4 - t2.8 you're not getting clinical, sharp footage. Thank the Maker.
@BRai4now yes, softer, but certainly there's a difference between softer film lenses of the 1980s, and phone cameras of the 2010s. The latter fails to attain the base detail. I was referring to the latter. Even if you put vaseline on a lens in the 80s, you still got grain at least, but I would argue that it also picked up texture details better than earlier digital cameras, even when softer.
@Jellai my Intention with the LoRA was to get rid of perfection, clean looks , plastic look or sweaty over tanned lock of the most babes. Also the lack of camera frames, filters and angles knowledge was frustrating. So I trained the LoRA. Over all i will improve and grow the dataset. 1 Thing I need to mention: my example videos are bad. I used a turbo LoRA for faster generation and I tested it without any turbo LoRA. What can I say way better quality results.
@Astroburner Yes. I never meant to say your lora was bad. Just that I needed to see comparison videos because the examples didn't let me know enough. I made the example videos, and I think they look great. I wasn't describing what your lora does to H3. I was describing why I had trouble seeing what the lora can do in the existing videos. I think my examples pulled more detail out, without needing new prompts, and I like the results a lot.
@Jellai why are you talking about that in the context of this lora or the examples posted. Completely irrelevant.
@BRai4now That they look like older smartphone videos? It was relevant to me, especially since it's a pain to get yellow buzz simply to test on my own setup. No longer relevant now that I can see that the lora can pull out more film-like quality out of H3, after testing on my own, except that people keep telling me that I'm too young, or don't know what films look like.
I'm not sure why my perspective is relevant to you, for you to interrogate it. My perspective is changed at this point, after 1) requesting comparison examples since yellow buzz is difficult, and I recognized that the original example videos likely weren't the best quality that could be pulled out (especially since they're named "Detail Enhancer"), 2) @Astroburner kindly donating the lora to me to test, 3) my having spent a chunk of today testing it under a variety of setups (so many 12 second 1MP gens), then posting the impressive results on civitai, and 4) conversations here. If you haven't been following that journey, you can't understand why anything is relevant to me at any point in time.
@Jellai you aren't a good enough judge if you think anything they posted looks like an old smartphone pic. Sorry. That's laughable. LOG profiles for creative grading didn't even exist back then. You couldn't get that kind of tonality or color.
I'll nope right out of any further discussion.
@BRai4now Yeah okay. Thank you for noping out. What a waste of time, dealing with you prodding at me for no reason. Have fun doing that to the next person who has a perspective on something.
I like what I feel like the lora is giving me. Something you might consider for another version is more training of close-up styles. H3, even with your lora, seems to only be able to do CUs and ECUs as if shot with a wide angle or short-tele lens and the camera moved closer to the subject. Invoking actual focal lengths for tele/portrait/close-up lenses like 65mm, 75mm, 100mm, seems to have zero effect and what you get is close-up distortion rather than the flattening effect of a proper close-up lens. I will offer, however, turbo appear to make the matter worse. I get a better ECU with no turbo and 25 stops than with the leading turbo setup but it's still got too much distortion for how a proper close-up lens would render a face if you weren't going for Terry Gilliam or Sam Raimi.
And regardless of prompting for it, successful anamorphic bokeh is random at best, if not entirely placebo or no more reliable than base model with no lora.
Also, H3 itself seems to be only clever enough to be dumber than if it wasn't clever at all with regards to DOF and overall scene lighting. Trying for major out of focus with daylit shots is difficult because it seems the training data might think that's more applicable to night shots, since there will be more footage of large f-stop usage at night but all it takes is heavy ND on the lens, letting you open up your stop even in bright scenes. Would be neat if you could prompt for ND and that sort of look.
Thank you for your great Feedback will adapt it in the next Version xD
@Astroburner Yeah, most of the limitations seem like base H3 being very versatile for wider footage but not so much the kind of lens selection that's more common for cinematic close-ups and object insert shots. I've seen successful "macro" really super close, but that's different.
It would be nice if you could do a close-up where the bokeh are bigger than maybe nickel or quarter size, and get some baseball and softball size circles of confusion with a flattering compression of depth. But I'm thinking specialized training imagery would have to be utilized showing bokeh and depth compression in a controlled way (ie. mannequin(s) or person(s) with a few LED back behind them in depth shot with various lenses at several apertures and focus distances). But that might be too much hassle, and just some training data with some really nice night and non-night footage showing big smooth bokeh will do.
For an example of what I mean, check this out from Blade (1998). Some great looking anamorphic photography in there. Here's daylight footage shot on long lenses with big f-stop to isolate him both in crowded scenes, and the gorgeous shots where he confronts Frost in the park.
https://www.youtube.com/watch?v=E4eb_ozkvXo
...in particular look how soft things get behind Blade in the ECU around 02:38 and the cut back to Frost. And then how great the slowmo section at 03:13 looks, which is clearly anamorphic but almost has a Petzval "swirly" bokeh quality that makes it extra dreamy.
请问之后会开放下载吗？Yellow Buzz太难获得了。
加入了三年都没有500
感谢您的辛勤付出
我给你发送 500 个 Yellow Buzz。这样你就可以购买我的 LoRA 了。祝你玩得开心。
Any plans for a public release down the road? Yellow Buzz is super hard to get, 3 years in and I’m still under 500 haha. Huge props for the amazing work though!
You can't buy it with blue? I bought it with blue
@luka0611 yes blue Buzz is easy to get, i accept for my Models also blue xD. Once i release v03 , V02 + v01 will be free.
@luka0611 Hi ! How you buy yellow buzz with blue ? A have so many blue but Yellow it's impossible tu buy with this wasfull crypto system
@Skeyz for all of my LoRAs i accept blue Buzz also. Can be bought with normal Credit Card on civitai.com
@Astroburner Dman you're right I didn't see the blue buzz button ! Let's go give you some blue buzz so :)
Unless you trained it on your own films, you're in the wrong profession for charging money.
I dident train it on films. The dataset is created by me with multiple handcrafted Pictures and later used some filters on top before training. XD
really want it, but I don't have any way to make a payment.
nice
Whoever said that this Lora doesn't work please shut the fuck up cause that's just skill issue. V2 just solve my problem, now the model understand blocking, staging, framing, cinematography terms without me having to tries 1000000x times to find a better grammar to do it.
Hello, could I ask how I can top up Yellow Buzz more easily? It’s a bit tricky. I really like your LoRA.
Ok friend have fun with my LoRA and my Buzz xD
@Astroburner Thank you so much, I really like your LoRA.
请问之后会开放下载吗？Yellow Buzz太难获得了。
辛苦了 V1.0 我就覺得很厲害了
我理解你，而且你知道嗎？我會幫你的。我會寄給你 500 個 Buzz，這樣你就可以用我的 LoRA 玩玩看了 :-)@Astroburner 優良作者必須支持!!!
Hey, do you have it available anywhere else where I can just pay with a normal credit card?
You can support me on patreon (https://www.patreon.com/c/astroburner) or on Ko-Fi (https://ko-fi.com/astroburner) , i send you soem buzz so you can have also fun with my LoRAs xD
@Astroburner Wow, thanks man.
@valentinospades402 Have fun with it dude xD
Where is this?
"separate prompt pack containing 15 complete H3 Text-to-Video prompts across film, games, romance, action, trailers, advertising and music-video scenarios."
on my patreon page