What actually runs on 8 GB of VRAM — six models, measured
RTX 3070 Ti Laptop, 8 GB VRAM, Ryzen 7 6800H, 32 GB RAM, Windows 11, ComfyUI 0.42.1 portable. No system tweaks, no --lowvram, nothing unloaded between runs except the model swap.
Every number below is measured on that machine with the same prompt, the same seed, three timed runs per configuration after one discarded warm-up, and the median reported. Resolution is 1024x1024 for all six so the comparison is fair — these models have different native sizes, but speed comparisons across different resolutions tell you nothing.
The numbers
Qwen-Image 2.1 — 4 steps, CFG 1.0, Lightning LoRA 1.0 — 16.0 s (16.1 / 16.0 / 16.0)
NoobAI-XL v1.1 — 28 steps, CFG 6.0, dpmpp_2m karras — 18.3 s (20.1 / 18.1 / 18.3)
MOSS 2.5D — 25 steps, CFG 2.5, er_sde simple — 38.2 s (38.2 / 38.1 / 40.2)
Qwen-Image 2.1 — 8 steps, CFG 2.0, Lightning LoRA 0.5 — 48.4 s (48.6 / 48.4 / 48.3)
Z-Image int8 — 30 steps, CFG 4.0, res_multistep — 58.4 s (58.4 / 58.4 / 58.5)
FLUX.1 dev Q4_K_S — 20 steps, guidance 3.5 — 72.5 s (72.5 / 72.4 / 72.5)
What this does not say
Faster is not better. NoobAI at 18 seconds is the fastest anime model here and it also has the worst hands — its 2.6B parameters show up exactly where you would expect, in fingers and small structures. Qwen-Image at 16 seconds in 4-step mode is fast, but at CFG 1.0 the negative prompt is ignored entirely, so you cannot suppress anything. The 48-second 8-step configuration exists because that is the cheapest setting where the negative prompt starts working.
Memory, and where it actually breaks
Peak VRAM as reported by nvidia-smi, including the desktop compositor which sits around 1.2 GB on this machine:
Qwen-Image 4-step — 7323 MB
Qwen-Image 8-step — 7671 MB
Z-Image int8 — 7675 MB
NoobAI-XL — 7553 MB
MOSS 2.5D — 7507 MB
FLUX.1 dev Q4 — 6983 MB
Nothing OOMs at 1024x1024. FLUX has the lowest peak because the text encoder gets offloaded to system RAM between the encode and the sampling pass — the 4.56 GB T5 does not live on the card. That offload is also why FLUX is the slowest of the six despite being the largest model: it moves weights back and forth instead of keeping them resident.
The failures start with resolution, not with the models. At 1024x1536 FLUX needs roughly 40% longer and sits much closer to the ceiling. The one configuration that reliably produces an out-of-memory on this card is stacking a ControlNet on top of a non-quantized checkpoint at 1536 pixels.
Portrait sizes, measured separately
The comparison above is square so the models are comparable. What people actually render is portrait, so here are the same models at the aspect ratios they are normally used at — single runs, same machine, not part of the controlled comparison:
Qwen-Image 4-step at 1024x1408 — 24 to 28 s
Qwen-Image 8-step at 1024x1408 — 72 to 104 s
NoobAI-XL at 832x1248 — 18 to 24 s
MOSS 2.5D at 1024x1536 — 66 to 72 s
FLUX.1 dev Q4 at 1024x1536 — 108 to 116 s
Practical picks
Fastest iteration: Qwen-Image 4-step. Sixteen seconds means you can throw away twenty drafts without thinking about it. Use it for composition hunting only.
Best hands and materials: Qwen-Image at 8 steps. Double the time of NoobAI, but fingers, fabric and specular highlights are in a different class.
Truest anime line work: NoobAI. It is the only one here that draws line art instead of rendering. Pair it with an OpenPose controlnet if the pose has to be exact — the text alone will not give you a difficult pose.
Two and a half D look: MOSS. Slower than NoobAI, softer than Qwen, and the one that sits between them stylistically.
If you need prompt adherence over style: FLUX. It follows complex spatial instructions — two people, a reflection, an object passed between hands — better than anything else on this card, and it is also the slowest here.
One correction to a common belief
The 4-step Lightning setup is usually presented as a straight upgrade: same model, four times faster. It is not the same model. Turning the LoRA back up to 1.0 and dropping to CFG 1.0 removes the negative prompt from the equation entirely, and the illustration quality changes — the 8-step output is what the model actually looks like when it is allowed to converge. If you are only making drafts, that does not matter. If you are making something to publish, it does.
All the workflows behind these numbers are published on this profile, free, with the models and settings listed.