Running FLUX image models locally on a Mac for video pictures
FLUX.2 klein 4B vs SDXL, on an M1 Pro
Every AI-picture story in TallyReel needs one picture per scene, and the same characters have to look the same from scene to scene. Studio draws those pictures on your own computer, so the image model has to be good, licensed for a paid app, and fast enough on a laptop.
On 30 September 2026 we replaced the two SDXL models Studio used with FLUX.2 klein 4B and tested both on the same five-scene story on an M1 Pro MacBook with 32 GB of memory. Here is what we measured and what we saw.
By Danish, who makes TallyReel ยท Published
What we found
FLUX.2 klein 4B clearly beat SDXL in all three art styles we tried, mainly because it follows long prompts: SDXL's 77-token limit cut off the character descriptions, so a character never appeared. In the shipped 4-bit MLX build, klein draws a scene in about 50 seconds on an M1 Pro and a whole five-scene video took 254 seconds. It is slower than SDXL's 30 seconds a scene, and worth it.
The test story
We wrote one five-scene story and drew it in three styles: storybook 3D, flat cartoon and soft anime. A small grey kitten called Mishi with a red knitted scarf carries chai through a bazaar, slips in a puddle, and an old tortoise shopkeeper in round brass glasses hands her a fresh cup. The last scene has both of them on a rooftop at sunset.
Two characters, a specific scarf, specific glasses: it is a simple story, and a hard test for consistency.
SDXL vs klein, same scene
The biggest difference came from prompt length. SDXL reads only the first 77 tokens of a prompt. Once the style, the scene and two character descriptions are in, the tortoise's description falls past that limit, so SDXL never drew him. It drew a second cat instead. FLUX.2 klein reads the whole prompt, so both characters appear as described.
Speed and memory on an M1 Pro
The first 8-bit timings made a 12-scene video take 25 to 30 minutes, which is too slow. Two changes fixed most of it. We moved to the 4-bit MLX build, about 4.6 GB to download, and drew at 544x960 before upscaling. Then we found that part of the slowness was memory pressure: limiting MLX's cache and clearing it between pictures brought times that had crept to 110 to 170 seconds back to about 50. A whole five-scene video, with voice, captions, music and the final render, took 254 seconds.
| Setup | Time per picture | Notes |
|---|---|---|
| SDXL (before) | About 30 seconds | Fast, but drops long prompts |
| klein 4B, 8-bit, prompt only | About 72 seconds | First test |
| klein 4B, 8-bit, one character reference | About 112 seconds | Peak memory about 13 to 14 GB |
| klein 4B, 8-bit, two character references | About 169 seconds | Slowest case |
| klein 4B, 4-bit MLX, as shipped | About 50 seconds | Drawn at 544x960 and upscaled |
Measured on an M1 Pro with 32 GB of memory, 30 September 2026.
Keeping characters the same: cast sheets
Prompt text alone does not keep a character identical across scenes. So before the scenes, Studio draws a cast sheet for each recurring character, one clear picture of them, and passes it to klein as a reference image when that character is in a scene. A cast sheet takes about 43 seconds. Studio works out who is in each scene from the script, so a scene only carries the references it needs.
Cast sheets are what made the side-by-side pictures above consistent. They are also where klein made its strangest mistakes.
Bugs we only found by looking
None of these showed up in logs or tests. We found each one by opening the pictures and looking at them.
- Twins: we first drew cast sheets with two views of each character. klein read that as two characters, so scenes came back with twins. Cast sheets are now a single pose.
- An extra companion: naming the characters before saying how many there were made klein add someone. Scene prompts now open with "One character only" or "Two characters only".
- Signatures: klein takes no negative prompt, and the painterly styles started signing their work. The prompt now asks for no signature or watermark.
- A duplicate: once, in the cartoon style, klein drew the tortoise twice in the same scene. In the anime style it once turned the tortoise into a man.
- Background extras: in crowded scenes an extra person sometimes appears. It depends on the random seed rather than the wording, so the fix is a Draw again button for that one scene.
Why klein, and not a bigger model
FLUX.2 klein 4B is released under Apache 2.0, which lets us ship it inside a paid app. Several stronger-looking models were ruled out by their licences, including the larger FLUX.2 dev and 9B models and Kontext. klein is also small enough for a laptop: the 4-bit build fits comfortably in 32 GB of memory alongside everything else a video needs.
On Windows and in the cloud the same model runs through diffusers on NVIDIA cards. On a cloud L4 GPU a picture takes about 3 seconds instead of 50, which is what makes AI-picture stories affordable in TallyReel's web and phone apps.
If you want to run it yourself
- On a Mac, use the 4-bit MLX build through mflux.
- Draw at a smaller size and upscale, rather than drawing at full 1080x1920.
- Cap MLX's memory cache and clear it between pictures, or later pictures get slower.
- Put the character count first in the prompt and ask for no signature or watermark.
- Look at every picture while you tune. The worst bugs will not show up anywhere else.
Models and links
Questions
Can FLUX.2 klein run on a Mac?
Yes. The 4-bit MLX build, about 4.6 GB, runs through mflux. In our test it drew a story scene in about 50 seconds on an M1 Pro with 32 GB of memory.
Is FLUX.2 klein better than SDXL?
For story pictures, clearly yes in our test. SDXL's 77-token prompt limit dropped character descriptions, so characters went missing. SDXL is faster, at about 30 seconds a picture.
Can FLUX.2 klein be used commercially?
FLUX.2 klein 4B is released under the Apache 2.0 licence. Always check the licence of the exact weights you download.
How do you keep the same character across scenes?
Draw one cast sheet picture per character first, then pass it to klein as a reference image in every scene that character appears in.
Does TallyReel draw AI pictures in the browser?
Yes. The web and phone apps draw them on cloud GPUs, and Studio draws them on your own computer.
Pick one. Make it tonight.
Type the idea into TallyReel, read the script, and it makes the video for you: in the browser, on your phone, or in Studio on your own computer. Start on the free plan.