Can a 16 GB GPU Finally Spell? (MOSS-Image vs FLUX 3 vs GPT & Gemini)

MOSS-Image is a brand-new open AI image model that claims it can spell and put every object exactly where you tell it. We ran it in ComfyUI on one home PC with a 16 GB card, against FLUX 3 Klein (also free) and two paid models, GPT image 2 and Gemini 3 Pro Image. Same prompts for all four. ⏱ CHAPTERS 0:00 Four models, one question 0:21 The line-up (and does it fit on 16 GB?) 0:50 Can it spell? Signs, posters, a menu, a letter 1:35 A four-panel comic (and the Director's catch) 1:53 Where it breaks: counting and the clock 2:19 The fix: give it a map 2:59 The character sheet we used to pay for 3:39 Action: punches, kicks and a fight 4:29 One face, four places 4:48 The scorecard 🧪 WHAT WE FOUND (our tests) • Text: a tie. A ten-line chalkboard menu with prices, a five-line poster, a 60-word handwritten letter: the free models kept up with the paid ones. • Counting and the time: the paid models win on their own. 100 numbered drawers: MOSS got about 53. Asked for 3:40, MOSS drew 2:37 and FLUX 3 drew 10:10. • Give it a map and it's fixed: layout boxes took MOSS to 96 of 100 drawers and exactly 20 people, and hands drawn in code gave 3:40 exactly. • Faces: FLUX 3 Klein kept our host's face every time (face match 0.91). MOSS gave four different women. • Speed on our card: FLUX 3 ~17 s a picture, MOSS ~60 s (a 4-step version: ~6 s). Cost: $0. GPT ~25¢ a picture, Gemini ~13¢. • Our pick: FLUX 3 for people, MOSS for signs, posters and layouts. 🛠 HOW WE TESTED • One RTX 4070 Ti SUPER (16 GB), 96 GB RAM, ComfyUI v0.43.0. • MOSS-Image: 25 steps, CFG 2.5, er_sde/simple, skip-layer guidance on layer 22, seeds 42 and 7. FLUX 3 Klein: 16 steps, CFG 1, seeds 42 and 7. • GPT image 2 (high) and Gemini 3 Pro Image (2K) through their APIs: one try per prompt, no retries. • The paid models are the bar: we tuned the free ones to catch up, not the other way round. • Drawers and people counted by hand at full size. Face match = insightface (antelopev2) similarity to our host's reference picture. • Every prompt, word for word, is in the comments. 🔗 MOSS-Image (OpenMOSS, Apache 2.0): https://huggingface.co/OpenMOSS-Team/... ComfyUI files: Comfy-Org/MOSS-Image · Edit LoRA: chinoll · 8-step LoRA: lodestones · 4-step LoRA: akhaliq · colour/depth LoRA: darask0 FLUX 3 Klein: Black Forest Labs Clio is an AI host. Her lab set was made with FLUX 3 Klein; her voice is Chatterbox and her lip-sync is LTX 2.5, all free and local. Every test picture is the named model's own output, mistakes included. This thumbnail was made by MOSS-Image too (from Clio’s picture), on the same 16 GB GPU. #MOSSImage #FLUX3 #LocalAI #16GBVRAM #ComfyUI