In front, 14 of your basics ComfyUI I got it: machine, graphic, style, control, face lock, zoom in and even video. But I know a lot of friends are stuck on one of the most realistic issues--There's not enough.
I was the victim of the 8G card. Watching the Flux show, full precision, all kinds of big models, it's like CUDA out of memory, and it's just falling apart. And then I realized, instead of being perfect, it's better to look for a model and a quantitative play that's "breathing the memory." It's about one I've been playing with latelyBoogu-ImageAND AN OPEN SOURCE MODEL THAT IS PARTICULARLY FRIENDLY FOR LOW-VISIBILITY STORAGE, AND 8G CARDS CAN RUN。
Note: This work stream is organized by official document + community, and the author himself did not make a hand-held cut-off on 4060 and did not measure the video. Parameters have been cross-checked with nodes, but please do a self-portrait; pit-scrap stories are summarised in the same model and are not recorded on the same day。
Let's get this straight
Boogu-Image 0.1 is an open-source image generation model, structure and OmniGen all the wayThe Chinese and English bilingual hints, and the layouts are very stableAnd officially, it's designed to take care of low visibility。
One, three variants
| Variable | What for | Savings |
| Base | Mass ceiling, Vincent's main power | Full accuracy is the best |
| Turbo | Distillation acceleration, step 4 | AT BEST, 6G CAN RUN |
| Edit | Reference Chart Edition (Chart) | Medium |
Remember:If you want to save the memory on Turbo+ Quantification, if you want quality, Base, if you want to change the map, Edit。
II. Documentation preparation (by mistake)
Low visibility GGF QUANTIFIED ROUTE(Small file, fit into 8G). Remember this. Wrong ComfyUI never reads:
ComfyUI/models/unet/ →boogu-turbo-dit-Q4_0.gguf
Qwen3vl_8b_fp8_scaled.safetensors
CofyUI/models/vae/ fflux1_vae_bf16.safetensors
CommyUI/models/loras/→ boogu_image_turbo_lora_rak_128_bf16.safetensors (optional, only Base/Edit, not Turbo)
if you're on the official original version of safetensors (non-quantifiable), the main model is placed in models/diffusion_models/, the file name is boogu_image_turbo_fp8_scaled.safetensors. the two routes are different, don't mix。
Trail experience:The first time I went to Boogu, I shoved the main model into the models/unet/ and used the original safetensor file name, the loader couldn't be cleaned for half an hour. GGUF will be on unet, in the original version, in pairs。
Two early dead pits:
① The ComfyUI version with PR #14523 must be updated(Natural support for boogu architecture). There are no boogu options in the old Load CLIP. Update: cd ComfyUI gill pip install-r updates.txt。
② The encoder can't be misusedI don't know. The standard SD/Flux encoder must be used with a specific VAE for Qwen3-VL FP8 encoder + Flux. 90% of the web-based "soft hair" feedback is from encoder or VAE。
iii. Hands to nodes
Boogu uses the native ComfyUI nodeNo need for that old plug-in(For old versions, no new ones)。
Load Diffusion Model (GGUF with UNnet Loader GGUF)
Load CLIP (type with boogu)
Load VAE, VAE Decode, Save Image
One sentence:Spread model into model mouth, CLIP with boogu into positive and negative hint, VAE to solve. It's almost exactly the same as Flux。
Critical pit:Load CLIP type must be created manually booguI don't know. Default may be a different structure, but it doesn't come out of a patch. For the first time, I thought the model was broken, and it was broken twice, and it was a night. Boogu's crip type is boogu, not sd3, not flux。
★ Practical closed ring: 10 minutes, with Turbo to produce a bilingual C-English (do now)
Preconditions:ComfyUI has been updated to include PR #14523 (updated by command in section II), otherwise there are no Boogu options in the Load CLIP。
Objectives:A Q4_0 quantitative version of Boogu-Image Turbo is used to produce a technical wind poster with Chinese titles。
Step (sight, don't jump):
1. Update ComfyUI:git pull + pip install-r requirements.txt, restart。
2. Down Model:unet with boogu-turbo-dit-Q4_0gguf, text_encoders with qwen3vl_8b_fp8_scaled,vae with flux1_vae_bf16, reset with second section directory, F5。
3. Add Node: Load Diffusion Model (GGUF with UNnet Loader GGUF)。
4. Connect: MODEL mouth of Diffusion Model → model mouth of KSampler ; CLIP mouth of Load CLIP → two clip mouths of CLIP Text Encode (positive/ne); Load VAE → Vae mouth of VAE Decode = latent → VAE Decode samples mouth of KSampler。
5. SET CLIP TYPE + SAMPLE: type for Load CLIP (this step is not in front of you); KSampler for Turbo Parameters - Steps 4, CFG 1.0, sampler euler, Scheduler sgm_uniform。
6. Write hints out: Writing "A technical launch poster with a headline on the NOVA computing platform, blue and purple glass formulation, three functional cards" (Turbo for Chinese, written directly in Chinese), point Queue。
Checkpoint
1 Normal: 4 steps and a few seconds, with a clear Chinese title on the poster。
2 Algorithm A = full-screen spatter/software: 80% of Load CLIP type does not have Boogu, back to section III。
3 Anomalous B = no model found with a load of error: The directory is misplaced, turn back to three positions in section II。
Job (selected):Move Steps from 4 to 8 and the CFG refers to 2.0, compares details to changes, and feels the Turbo step number of sweets。
IV. CORE THREE EXCHANGE
| Variable | CFG | Steps | Sampler | Scheduler |
| Base | 4.0 | 30~50 | euler | simple |
| Turbo | 1.0 | 4 | euler | sgm_uniform |
| Edit | 3.5 | 25 | euler | simple |
My habits: daily drawings are brainless, Turbo (4 steps out), fine posters/products to cut Base to 40。You have to write long notesBoogu doesn't eat the short word "a cat" and the longer the more concrete the better, which is contrary to Flux. I just started taking Boogu, and when SDXL used it, the hint was "a cat," and I came out with an abstract painting; it went into three lines to describe it in detail。
V. Common pits at once
1. Load CLIP type does not have bogus Well, it's the most common。
2. ComfyUI is too old, no PR #14523 → The Load CLIP does not have a boogu option, update first。
3. ENCODER/VAE USED WRONGLY A moist hair with a Qwen3-VL FP8 encoder + Flux VAE。
4. Quantified Profile Error → Boogu has only Frat Quants (Q4_0/Q4_1/Q5_0/Q5_ 1/Q8_0)NO Q2_K/Q3_K_M THIS KQQDon't search under Q3。
5. Q5_1 CAN'T FIT 8G → Q5_1 ABOUT 8.6G, OVER 8G CARD LIMIT, 8G Q4_0 (APPROXIMATELY 6.8G)。
VI. QUIREMENTS OF PARLIAMENTS
| take | Variable | Quantified | Steps | CFG |
| 8G FAST WATER TEST | Turbo | Q4_0 (~6.8G) | 4 | 1.0 |
| THE 8G CARD IS GOOD | Turbo | Q5_0 (~8.0G) | 4 | 1.0 |
| 12G QUALITY PARTY | Base | fp8_scaled | 40 | 4.0 |
| Reference Figure | Edit | Q4_0 | 25 | 3.5 |
| 16G+ FULL QUALITY | Base | bf16 | 50 | 4.0 |
VII. 4060 / 8G SPECIAL
4060's 8G running BooguQ4_0 IS DESSERT(APPROXIMATELY 6.8G MAIN MODEL + ENCODER + VAE, TOTAL OCCUPANCY PRESSURE WITHIN 8G). POINTS:
· Quantified: THE MAIN MODEL USES Q4_0 GGUF (APPROXIMATELY 6.8G), WITHOUT TOUCHING Q5_1 (8.6G OVER), Q8_0 (11.6G DREAMING)。
· ENCODER WITH FP8:qwen3vl_8b_fp8_scaled, much less than full accuracy。
· Lower Display Mode: Start –lowvram or settings to open Low VRAM and then player offload if necessary。
· Respond to greed: 1k first, more than 2k directly, with the magnification of 13k。
Trail experience:I HAD A GREEDY Q5_1 THOUGHT, "QUALITY BETTER," AND AS A RESULT, 8G CARDS WENT STRAIGHT TO CUDA OOM AND EVEN THE INTERFACE GOT STUCK. BACK TO Q4_0 SECONDS OUT. THE LOW-VISIBILITY CARD TAKES THE LIFE OF Q4_0, AND THE QUALITY IS SUPPLEMENTED BY A HINT AND LATER MAGNIFICATION, SO THAT IT IS NOT AS STRONG AS THE QUANTITATIVE FILE。