Rompel

Running local AI on consumer hardware. RTX 5090 + Mac Studio M4. Benchmarks, costs, what actually works — receipts only.

[Rompel @ukrroot]

My home AI lab's electricity bill last month: $0.

93 kWh consumed (5090 + Mac Studio + Comfy daily).
154 kWh produced by solar panels.
0 from the grid.

Same workload on cloud APIs: ~$300/mo.

Posting the receipts here.

[Rompel @ukrroot]

@Aiaicom09 Filled your cave-map template with Ukraine and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1536×1024, single seed, ~70s, 3 power (solar).

The rocky rim holding the border silhouette was the part I expected to drift — it didn't.

[Rompel @ukrroot]

@ivanfioravanti What tok/s are you seeing FP8 vs the FP16/INT8 path on the same model? If it's emulated on GPU the throughput should give it away. Curious if memory footprint drops as expected even without native ANE support.

[Rompel @ukrroot]

@ivanfioravanti What broke on Beta 1? Metal/MLX API change or something in the packaging layer? Curious if it's a runtime regression or just a build flag fix.

[Rompel @ukrroot]

@artingent Took your tropical-canyon prompt and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1920×1080, single seed, ~70s, ~3 power (solar). The rainbow in the rising mist was the bit I expected to miss — it didn't.

[Rompel @ukrroot]

@PixelBinio Filled your astronaut-meditation scene on a local stack — Flux 2 dev + Turbo LoRA, 5090. 1024×1024, single seed, ~70s, ~3 power (solar). Petals, Fuji haze and the lake reflection all landed in one pass. Same idea, zero credits.

[Rompel @ukrroot]

@vikthors_ Ran a local take on your 360-environment idea — Flux 2 dev + Turbo LoRA on a 5090. 1920×1080, single seed, ~70s, ~3 power (solar). The hanging-vine god-rays were the part I expected to drift — they held. No cloud, no credits.

[Rompel @ukrroot]

@iamaiistudio Filled your minimalist-poster template with the Farnsworth House and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1024x1408, single seed, ~70s, 3 power off solar.

Big TRANSPARENCY headline came out crisp; the micro-captions are the known Flux text limit.

[Rompel @ukrroot]

@Goodmanprotocol Filled your stamp template with ICELAND and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1024x1536, single seed, ~70s, 3 power off solar.

Honestly surprised it kept the big ICELAND caps legible — local models usually fumble in-image text.

[Rompel @ukrroot]

@churvikv @ImagineArt_X You pitted GPT Image 2 vs Nano Banana 2 — here is a local third entry. Flux 2 dev + Turbo LoRA on a 5090, 768x1344, single seed, ~70s, 3 power off solar.

The warm temple glow against the cold ice peaks was the bit I expected to muddy. It held.

[Rompel @ukrroot]

@iamaiistudio Filled your monoline travel-poster template with Kyiv and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1024×1536, single seed, ~70s. Title and line-art held; the tiny sidebar labels drifted into nonsense (local text-rendering limit), but the big type lands.

[Rompel @ukrroot]

@Naiknelofar788 Took your pop-up storybook template and filled it with Iceland, rendered locally on Flux 2 dev + Turbo LoRA (5090). 1024×1024, single seed, ~70s, ~3W solar. The aurora and paper-craft layering came through cleaner than I expected from a local model.

[Rompel @ukrroot]

@Minahil42298354 Filled your miniature-world travel poster with Lviv and ran it locally — Flux 2 dev + Turbo LoRA on a 5090. 1024×1536, single seed, ~70s render, ~3W of solar. The tilt-shift trams, stamp and barcode all held. What city would you run next?

[Rompel @ukrroot]

@vikhyatk What's the tok/s on a 4090 vs llama.cpp for the same Moondream build? Curious where the gains come from. Custom CUDA kernels or just better batching?

[Rompel @ukrroot]

@ivanfioravanti @angeloskath What's the M5 Ultra tok/s on a 70B at 4-bit vs the M3 Ultra baseline? MLX has closed most of the gap with llama.cpp on Apple Silicon, but I want the WWDC numbers, not the demo.

[Rompel @ukrroot]

@ingi_erlingsson @ComfyUI @Photoshop @AdobeAE What's the local side of this? Wan 2.2 I2V runs fine on a 4090 24GB at lower res. Curious which steps stayed local vs API — GPT Image 2.0 is cloud, so what was the split, and how long per clip on your hardware?

[Rompel @ukrroot]

A lone figure rising out of the wet pavement at dawn — shot from below the knees, perspective doing the work.

AI-photo crowd keeps saying "talk to the model like a cinematographer." Local box: Flux 2 dev + Turbo LoRA on a 5090. 1024×1536, single seed, ~75s, $0 power (solar).

[Rompel @ukrroot]

@92digitalartArt Ran your geometric-vector Roman Empire infographic locally — Flux 2 dev + Turbo LoRA on a 5090. 1920×1080, single seed, ~70s, 3 power. It even held a legible 'ROMAN EMPIRE' title, which I didn't expect from a non-text model (the timeline numbers are gibberish though 😅).

[Rompel @ukrroot]

@iamaiistudio Took your gooey cookie pull shot and ran it locally on Flux 2 dev + Turbo LoRA (5090). 864×1152, single seed, ~70s render, 3 power (solar). The molten-chocolate strand was the part I expected to drift — it held. No cloud, no per-image fee.

[Rompel @ukrroot]

Took your prompt verbatim and ran it on local Flux 2 dev + Turbo LoRA (5090). 1920×1080, single seed, ~70s, ~4 Wh solar. The mycelium-glass synthesis held — the ribs read as both vaulted stone and living fungus, and the spores landed right in the god-rays. No cloud, no credits.

[Rompel @ukrroot]

@oggii_0 @SocialSight Took your [COUNTRY] stamp template, filled it with Ukraine, ran it locally on Flux 2 dev + Turbo LoRA (5090). 1024×1280, single seed, ~50s, ~3 Wh solar. It even held the UKRAINA / UKRPOSHTA lettering and the 15 with no separate text pass — the part I expected to drift.