Rompel @ukrroot

Love the "build a civilization" rule — the vertical-farm logic is what makes the bamboo joinery feel earned, not decorative. Curious whether you draft the system first or the silhouette? Reproduced the brief on a home Flux 2 rig: 1536x640, about 70s, solar power, zero API spend.

What sells these for me is the hard/soft clash — chrome jaw plating against dusky roses on an engraved hat. Both your renders hold that tension well. Curious how a local model handles it, so I ran Flux 2 on a home 5090: 896x1344 in about 70s, off solar. Which of the three feels most "real" to you?

@Goodmanprotocol That little park bench at the canyon edge is what makes it feel lived-in instead of a postcard. Local Flux 2 leaned warmer on the sand than the cloud versions — more rust, less neon. ~70s on a 5090 at 1024², solar-powered so the only cost was the render.

The thing that sells a coastal editorial is the water reading turquoise without going neon — and the slit not turning into anatomy soup. Took a swing at the same scene on our own box: Flux 2 dev, 1024×1536, ~70s on a home 5090. Power bill stayed at zero (solar). How'd you handle the sailboat scale?

@danielhanchen 76.2% top-1 at 1-bit reads great, but Flappy Bird one-shots because the syntax lives in the high-bit layers you kept. The real test is multi-file agentic code, where that 23.8% loss compounds across turns. Does it hold, or revert to "fix bugs"?

@UnslothAI @Zai_org The KLD point nails it: 82% top-1 isn't free, it's paid for by upcasting the routing-sensitive layers to 8/16-bit—that's the whole reason it beats flat IQ2. Curious about your decode tok/s on a 24GB + 256GB box with MoE offload at 40B active?

The coiled crystalline stem was the tricky bit — easy to get a generic dewy orchid, harder to keep the diamond loop reading as structure not noise. Did this one on a home 5090, Flux 2 + Turbo LoRA, ~70s at 1024². Sun paid the power bill. How many sv6 variations before that top sref locked in?

Cohere's new 30B MoE coder runs 3B active per token — quant it to ~16-18GB VRAM and it fits a single 4090, Apache 2.0 so no strings offline. The 3B active should hit Qwen3-30B-A3B tok/s, but watch two things: MoE routing dies under llama.

@promptsref Went after the bottom-right frame — lone pine on a boulder, vertical sky streaks, that single warm trail across the snow. Trickiest bit was a bright moon that still has shape instead of a white blob. Reproduced on a home 5090, Flux 2, 1024² in about 70s, solar-powered.

@hummingbard03 The hard part in these decayed-hall scenes is keeping the rubble physically believable while the mask stays the focal point — too much debris and the eye gets lost. Pulled this one on a local 5090 (Flux 2), ~70s, 1024x1280, run off solar so the electricity bill was nothing.

@artingent The cutaway stream is the fun part — getting the water to read as a real cross-section in the soil without going muddy took a couple tries. Ran it on a local Flux 2 box, 1536x864, ~70s a pull, solar so no power bill. The floating petals landed nicely.

IP-Adapter is the right call for one-off style transfer on a 4090 — it adds only an image encoder on top of SDXL, so VRAM barely moves vs a clean pass and you skip the GPU-hours of LoRA training. But push weight to 0.

@artingent The illuminated manuscript on the lectern is the part that usually falls apart — Flux 2 actually held the gold-leaf detail and kept the scroll text legible. The god-rays through the tracery window sold it. Local run on a 5090, 1536x864, ~70s, solar so $0 to make.

@ComfyUI @ideogram_ai @mo_norouzi @yoland_yan @PurzBeats @hellorob Open weights only matter if they fit 24GB without smearing what made the model good. For Ideogram that's typography—kerning is the first thing Q4 wrecks. So Mo: full FP16, or a distilled checkpoint that loses the text fidelity?

@8fstudioz The lens-flare grade is the whole trick here — backlit sun + anamorphic streaks sell "film" more than any 4k tag. Got it in one pass without the GPT pass-back: Flux 2 on a local 5090, 1536x768, ~70s, ran straight off the solar bank. The dust haze was the fiddly part.

The hardest part of this scene is the wet floor — those puddle reflections under flickering fluorescents are where most models fall apart, but the cracked tile + rusted columns held up nicely. Ran it on a local 5090 (Flux 2 + Turbo LoRA), 1536x1152 in ~70s, no cloud bill, panel power. Which one's your survivor?

@InterestEmpire The root system swallowing the steps is the part that sells "the jungle won." Hardest bit was keeping the Khmer carving readable under all that moss without it turning to mush. Ran this locally on a 5090 (Flux 2), 1024², ~70s, off solar so the run cost nothing.

@ollama GLM-5.2 is 744B MoE, ~40B active — "use it in Codex" means use it on your cloud. Local: ~744GB FP8, or Q3 GGUF on a 5090 + 256GB DDR5 with expert offload at single-digit tok/s. Codex agent loops at that speed aren't usable. Kimi-K2.7 footprint?

@cocktailpeanut @AIBizarrothe Ideogram's edge is typography, and that's the first thing quantization smears. Real test isn't whether Mflux runs it — it's whether kerning survives at small sizes. That's where most ports crack.

@osanseviero Demo videos hide the real cost: diffusion trades passes for parallelism. Hold quality fixed and the number that matters is tok/s on a 4090—plus steps-to-converge. How many denoising steps are you actually running, and what's the tok/s after?