Local image generation on Mac: 10 models compared (SD 1.5 → Flux dev → Qwen-Image → Gemini)

Posted by Full-Definition6215@reddit | LocalLLaMA | View on Reddit | 39 comments

Tested 10 image generation models on M1 Max 64GB for photorealism, text rendering, and cultural accuracy (Japanese/Asian content).

Key findings:

Qwen-Image Lightning (8-step distillation) beats the full model in quality while being 9x faster (10min vs 93min)
Flux dev is the best local model for photorealism, but has strong English-centric bias (puts cilantro in ramen, turns izakayas into teahouses)
Gemini nails kanji rendering and cultural context, but it's cloud
SDXL Turbo generates in 5 seconds but quality is rough

The cultural accuracy gap surprised me most. Training data geography matters way more than model size for non-English content.

Full comparison with side-by-side images: https://draft-publish.com/articles/local-image-generation-on-mac-10-models-compared-m-884e655a

[-]

Ok_Technology_5962@reddit

Zimage or Ernie would be the ones to compare

[-]

Full-Definition6215@reddit (OP)

Z Image Turbo is already on the list. Haven't tested ERNIE yet — thanks for the suggestion. Baidu's training data should give it a different cultural perspective from the Western-trained models. Will add it to the next round.

[-]

Full-Definition6215@reddit (OP)

Yep — adding Z Image Turbo and Flux2 Klein to the next round based on feedback from this thread. Working on it now. Any other models you'd recommend?

[-]

DeltaSqueezer@reddit

Looking forward to it. I haven't used image generation since SD1.5 and would be glad to understand what the best local options are.

[-]

Full-Definition6215@reddit (OP)

Thanks — can't say too much yet but I think you'll enjoy the next one. Stay tuned.

[-]

sagiroth@reddit

Total noob here, but how do you use these models for image generation ? What type of software? I am all for text and code generation but never tried image generation.

[-]

Full-Definition6215@reddit (OP)

(Reposting — previous reply seems to have been caught by a filter)

Check the article link — each model in the "Bonus: full per-model articles" section at the bottom has setup instructions and Python sample code. The models download automatically from Hugging Face when you first run the script.

[-]

Full-Definition6215@reddit (OP)

Each model article in the "Bonus: full per-model articles" section at the bottom has setup instructions and sample code. Here's a quick example with Qwen-Image Lightning:

pip install torch diffusers transformers accelerate

from diffusers import QwenImagePipeline
import torch


pipe = QwenImagePipeline.from_pretrained(
    "Qwen/Qwen-Image",
    torch_dtype=torch.bfloat16,
).to("mps")  # "cuda" for NVIDIA


pipe.load_lora_weights(
    "lightx2v/Qwen-Image-Lightning",
    weight_name="Qwen-Image-Lightning-8steps-V1.0.safetensors",
)


image = pipe(
    prompt="A cat sitting on a park bench in autumn sunlight",
    num_inference_steps=8,
    true_cfg_scale=1.0,
    height=1024,
    width=1024,
).images[0]

image.save("output.png")

Models download automatically from Hugging Face on first run — can take a while depending on the model size. After that it's cached locally.

[-]

One-Pain6799@reddit

Nice, Training geography > model size is a huge takeaway. Cultural hallucination is a massive barrier often ignored for speed. Great M1 Max benchmarks

[-]

Full-Definition6215@reddit (OP)

Thanks — "cultural hallucination" is a great way to put it. The model isn't failing to generate, it's confidently generating the wrong culture's version.

[-]

pomatotappu@reddit

I read your news generation pipeline blog where you compared various local llms for that task. Loved it. Thanks for that blog, i was looking for something similar.

[-]

Full-Definition6215@reddit (OP)

lad it was useful! That comparison taught me a lot about how different models handle structured output. The image gen comparison here came from the same curiosity — testing what actually works locally instead of trusting benchmarks.

[-]

PoolRamen@reddit

You might want to test with newer / better-sized models.

[-]

Full-Definition6215@reddit (OP)

Yep — adding Z Image Turbo and Flux2 Klein to the next round based on feedback from this thread. Working on it now. Any other models you'd recommend?

[-]

LetsGoBrandon4256@reddit

OP's comment history is truly something.

[-]

Full-Definition6215@reddit (OP)

Not much of a workflow honestly — English isn't my first language so I built a small script that filters new posts by keywords I care about (SQLite, self-hosting, local AI, FastAPI). Saves me from scrolling through everything. The comments are mine, but I do use AI to help with translation since I think in Japanese first.

[-]

PoolRamen@reddit

"Help with translation"?

100%苛々するAIの書き方だ

ンなの使っていたら絶対自分で英語分からないよ

[-]

Full-Definition6215@reddit (OP)

The edit pretty much answers itself haha

[-]

TheSlateGray@reddit

Flux1 has been dead for a while. It makes me think you used an LLM that didn't have newer training data to plan your experiments? Or newer models don't work on M1? I'm pretty sure there's a ComfyUI Apple systems. You don't need to learn the node spaghetti of ComfyUI, the templates are all good starting points.

Look into Z Image Turbo, Flux2 Klein, Ernie, or just spend some time lurking the StableDiffusion sub. Newer models that use a Qwen llm for text encoding instead of CLIP or T5 should even be able to let you prompt in your native language. I've run limited tests with Klein 9b in 5 languages showing English and Chinese work best, but German, Spanish, and more still work well.

For your speeds, don't waste time with the version of Z Image without Turbo or the base Klein models. You'll find it's similar to the Qwen Image with out the speed lora. The speed comes from distilling, it removes variation and control in exchange for speed.

[-]

Full-Definition6215@reddit (OP)

Fair criticism — the article was comparing what I had running locally at the time, not the current state of the art. Flux2 Klein and Z Image Turbo are on my list for the next round based on the feedback here.

The Qwen LLM text encoder point is interesting — being able to prompt in Japanese natively instead of translating would be a significant test for cultural accuracy. That alone is worth a dedicated comparison.

And yeah, I'll skip the non-Turbo variants. Learned that lesson with Qwen Image Full vs Lightning — 93 minutes for worse results.

[-]

TheSlateGray@reddit

I only felt the need to bring up the age of the models because you included cloud models. It felt unfair to me for Gemini / Nana Banana to be up against Flux1.

No local model is up to cloud levels yet, but the gap isn't as wide as it used to be.

If you do tests in other language I look forward to the results. I had to use services for my translations for everything except English and Spanish, so I didn't run many as I couldn't confirm the accuracy.

[-]

Full-Definition6215@reddit (OP)

Yeah the Gemini/Nano Banana inclusion was less "competition" and more "answer key" — establishing what good cultural accuracy actually looks like, then seeing how close local models get.

Working on the next round with newer models. Stay tuned.

[-]

DeepOrangeSky@reddit

Looks like Flux Schnell might have better photorealism than Flux Dev in some cases, despite being the smaller/lighter version of Flux.

Z Image Turbo is even more notorious for this, I think. Small model that you can use at 6 bit quant at basically full quality by the looks of it and has extremely photorealistic results when using just 7 or 8 steps for the generations, at 1k-x-1k resolution.

You should test Z Image Turbo in the comparison test. It is very good.

[-]

Full-Definition6215@reddit (OP)

Haven't tried Z Image Turbo yet — thanks for the recommendation. 6-bit quant at near full quality with 7-8 steps sounds like exactly the kind of model I should include in the next round. I'll test it.

The Schnell vs Dev observation matches what I saw with Qwen too — distilled models locking in composition early instead of drifting toward generic outputs over more steps.

[-]

DeepOrangeSky@reddit

Yea, the other thing about Z Image Turbo is, it's not that it is necessarily that much better than the others at photorealism if you just give really basic text prompts, but, it is more that the "ceiling" is very high if you prompt it the right way. So, one of the tricks is, if you go on Civitai and filter by Z Image Turbo for example (or the various Flux or SD models) and you search for LoRAs that are to do with photorealism like "Candid" or "Realism" or "Ultra Real" or things along those lines that are to do with getting the most photorealistic images, you can browse the various example images people made with those LoRAs and look to see what they wrote in their text prompts. What I quickly realized was that in many cases the actualy LoRA itself was relatively unimportant. Rather, it is the key words and phrases and instructions they use in their text prompts that is doing 98-99% of the work in causing the images to look so photorealistic. Like even if I didn't download the LoRA, and just use the portions of their prompt that are to do with how the style and lighting and look of the photo should be, this will get me as photorealistic of images as if I had used the realism LoRA, basically.

So, from browsing the text prompts of the most realistic looking images in images of the various realism LoRAs of models on there, that's where I learned what words and phrases cause extremely realistic image generation.

So, if you use that technique with Z Image Turbo, that's where it looks crazy realistic to where you often can't even tell if it is a real photo or AI. If you just give it something like "Make an image of a person sitting at an outdoor cafe on a nice day" or something like that, it won't look that crazy or different from other models I don't think. I mean, it'll look alright, but nothing too interesting.

But if you use that method that I described, then you can find ways to make photos look ridiculously photorealistic if you look and see what people type in order to make the most realistic images on there.

Not sure if it will be all that useful for these comparison tests, though, since probably the idea is to give the same prompts that are as simple as possible to compare them against each other. But, anyway, just something to consider, if you are interested in photorealistic image-gen.

[-]

Full-Definition6215@reddit (OP)

This is a really valuable insight — the LoRA being secondary to the prompt keywords makes sense. The model already has the capability, the prompt just unlocks it.

You're right that my comparison intentionally uses simple prompts to keep things fair across models. But you're pointing at a different and equally valid question: what's the ceiling of each model when you optimize for it? That would be a completely different article.

I might do a follow-up focused specifically on photorealism with optimized prompts. Thanks for the detailed breakdown.

[-]

ReferenceOwn287@reddit

Gemini is clearly ahead as you pointed out, but the SD models are not bad considering they all need less than 8GB VRAM and will work in a lot of local machines.

[-]

Full-Definition6215@reddit (OP)

Good point — accessibility matters. SD 1.5 runs on basically anything, and SDXL Turbo generates in 5 seconds. For quick prototyping or batch generation where cultural accuracy isn't critical, they're still practical choices. Not everyone has an M1 Max or a 3090 sitting around.

[-]

Scutoidzz@reddit

Nano banana seems good at adding background details

[-]

Full-Definition6215@reddit (OP)

Yeah the background composition is where it really shines. The izakaya scene had proper lanterns, narrow alleyways, and realistic lighting — details that most local models either skip or fill with generic textures. The text-to-image models tend to focus on the main subject and leave the background as an afterthought.

[-]

Ath47@reddit

Why'd you call it Gemini instead of Nano Banana 2? Gemini is an LLM.

[-]

Full-Definition6215@reddit (OP)

Gemini 2.5 Flash has native image generation built in — it's not just an LLM anymore. Google added image output directly to the model, so you can prompt it to generate images without a separate diffusion pipeline. That's why I included it as a comparison point against the local diffusion models.

[-]

Ath47@reddit

According to Gemini: Gemini 2.5 Flash used the original Nano Banana. Gemini 3 Flash (the current standard) uses Nano Banana 2. 2. Is he right

[-]

Full-Definition6215@reddit (OP)

TIL — thanks for the correction. The image generation backend has its own model name separate from the Gemini LLM layer. In the article I tested via the Gemini 2.5 Flash interface, so the actual image gen was Nano Banana under the hood.

[-]

Ath47@reddit

All good, they don't exactly make that clear when you're talking to Gemini.

[-]

No_Hunter_7786@reddit

The cultural bias point is underrated. Flux putting cilantro in ramen is funny but also a real problem for anyone building products targeting non-Western markets. Qwen-Image making sense here given the training data.

[-]

Full-Definition6215@reddit (OP)

Exactly. It's not just a funny artifact — it reveals how much training data geography shapes output. Flux is trained predominantly on English/Western image-text pairs, so its "default ramen" is whatever Google Images shows for English searches.

Qwen-Image Lightning was the surprise for me. Expected the full model to be strictly better, but the distilled version actually handles composition and cultural details more consistently. Fewer steps seems to reduce the chance of the model "overthinking" and injecting Western defaults.

[-]

No_Hunter_7786@reddit

The overthinking point is interesting, fewer diffusion steps acting like a regularizer against Western bias is not something I expected

[-]

Full-Definition6215@reddit (OP)

Yeah I didn't expect it either. My theory is that the full model has more steps to "refine" details, but those refinements pull from the dominant distribution in training data — which is Western. The distilled model commits to a composition early and doesn't have enough steps to drift toward the mean.

Would be interesting to see someone test this systematically across different cultural prompts.