Source: raw/Quantum_breakthrough_Gemini_4_Argon_MOSS_Image_KLING_4_OpenAI_spy_Muse_Wonder_-_AI_NEWS.md (AI Search news, 2026-10-18, for MOSS’s licence) and raw/New_top_local_AI_image_generator_is_here.md (AI Search, MOSS Image review and tutorial, https://www.youtube.com/watch?v=cUeZMUGCxrw, uploaded 2026-10-20, about 22 min; sponsored by Higgsfield) and raw/Finally_New_best_local_AI_image_editor_is_here.md (AI Search, Qwen Image 2.1 review and tutorial, https://www.youtube.com/watch?v=BaE6UBfNdQk, uploaded 2026-09-22, about 20 min; sponsored by Luma). Platform: YouTube auto-captions, fetched 2026-10-30. Caption fixes: “Quen / Quinn / Kquen” → Qwen, “Crea 2 / creatu” → Krea 2, “ideoggram” → Ideogram, “Chromos” → Kromoss, “Ostracus AI toolkit” → ostris ai-toolkit, “Abray” → Abiray, “coin 3.5 9B” → Qwen 3.5 9B.
Two free, open image models that run offline in ComfyUI on modest GPUs, covered as step-by-step install tutorials. MOSS Image is a 5B text-to-image model, small enough for about 4 GB of VRAM, with strong built-in knowledge, good in-image text and bounding-box layout control. Qwen Image 2.1 (Alibaba) generates and edits: natural-language edits with up to 10 reference images, and images with a transparent alpha channel. Both claims of “best in class” are the presenter’s opinion from his own tests; the generation times are from his laptop’s RTX 5000 Ada (16 GB).
Key Takeaways
- MOSS Image is Apache 2.0 per the presenter’s news episode, so commercial use is allowed; Qwen Image 2.1’s model licence is non-commercial (see below).
- Pick by job. MOSS for fast, cheap text-to-image on small GPUs; Qwen Image 2.1 when you need to edit existing images in plain language or need transparent PNGs. MOSS cannot do semantic edits, only classic masked inpainting.
- Both install the same way: update ComfyUI, open its built-in template for the model, download the diffusion model, text encoder and VAE into the matching
models/subfolders, press R to refresh, and select them in the loader nodes. - Choose the quantization by VRAM. MOSS: BF16 (10 GB) if you have more than 8 GB, otherwise the INT8 build. Qwen Image 2.1: BF16 is 14.2 GB, INT8 is 7.2 GB (fits about 8 GB), and community GGUFs go down to 3.19 GB for 4 GB cards.
- Speed on his 16 GB laptop GPU: MOSS about 25–30 s per image, or three to four times faster with an 8-step DMD LoRA. Qwen Image 2.1: 29 s text-to-image, 81 s for an edit, 116 s for background removal, 128 s for an edit with a LoRA at 2 MP.
- MOSS against its rivals (his tests): Krea 2 is better but “more than twice the size”; MOSS beats Z-Image (6B) on understanding and flexibility. Fine-tunes such as Kromoss already look better than the base model; he calls it “the SDXL era all over again.”
- Qwen Image 2.1’s licence bars commercial use of the model without a separate licence, but per Qwen’s reply he relays, generated images are not licensed materials, so you keep the rights to your outputs. Verify before client work.
- Skip Qwen’s prompt-enhancer files on low VRAM. They are about 9.5 GB each; he advises writing a good prompt by hand instead.
MOSS Image (5B text-to-image)
What it does well (his demos): realistic photos, anime and art styles, very wide or tall aspect ratios, celebrities, anime and game characters, brand names and car models known “right out of the box” with no LoRA, and dense text: diagrams and infographics.
Two model variants: SFT for generating images, Base for training and fine-tuning. Use SFT unless you are training.
Install (ComfyUI, Windows portable):
- Run
update/update_comfyui.bat, then start ComfyUI. - Templates → search “moss” → the MOSS Image text-to-image workflow.
- Download from
huggingface.co/Comfy-Org/MOSS-Image:- diffusion model (SFT BF16 at 10 GB, or the smaller INT8 build) →
models/diffusion_models - text encoder (Qwen 3.5 9B: BF16, INT8 or the smallest W4A8) →
models/text_encoders - VAE (full or the 250 MB small one) →
models/vae
- diffusion model (SFT BF16 at 10 GB, or the smaller INT8 build) →
- Press R, select the three models, set aspect ratio and megapixels (1 MP square is about 1024×1024), write a prompt, optional negative prompt, and run.
Settings he explains: the prompt-enhance toggle, seed (same seed and settings reproduce an image), steps (little gain past 40–50), CFG (raise it if the image ignores your prompt), sampler and scheduler.
Layout control. The layout-guided template lets you draw bounding boxes on the canvas and caption each one (“red sofa”, “a window with a T-Rex outside”), plus a whole-image description. His test placed every boxed element where drawn. If the box editor does not show, enable Nodes 2.0 in Settings.
LoRAs. Inside the workflow’s subgraph, put a Load LoRA node before the KSampler’s model input, pick the LoRA, set strength, and add its trigger words to the prompt. Available LoRAs he shows include anime styles, realistic skin and an ink style.
Faster: 8-step DMD LoRA. Load lodestones/MOSS-Image-SFT-8-steps-dmd at strength 1, set steps to 8 and CFG to 1. At CFG 1 the negative prompt has no effect; raise CFG slightly if you need it.
Better quality: fine-tuned checkpoints. Download a community checkpoint such as Kromoss (lodestones/Kromoss, BF16 or INT8) into models/diffusion_models and select it in place of the base model.
Editing workaround: inpainting. Add a Load Image node, paint a mask in the image editor, route it through VAE Encode (for Inpainting) into the KSampler’s latent input, bypass the empty-latent node (Ctrl+B) so the output uses the input image’s size, and prompt for what goes in the mask. “It’s not great… This is just traditional inpainting.”
Training your own LoRAs: ostris ai-toolkit (github.com/ostris/ai-toolkit) already supports MOSS Image and runs on macOS.
Qwen Image 2.1 (Alibaba, generate and edit)
What it does (his demos): realistic photos, good anatomy and prompt adherence, complex infographics and UI designs, and:
- Transparency. Outputs with an alpha channel, and edits that keep it (change the text on a transparent graphic; extract foreground leaves and branches as their own transparent layer).
- Multi-reference editing, up to 10 images. Combine a person, a jacket, shoes, a handbag and a hat from separate photos into one image, or furnish a room from product shots, with facial and character consistency.
- Annotation edits. Draw on the photo over items to change or remove, or where something should be added.
- Outfits and panoramas. Put a complex outfit on a model while keeping its design; turn a selfie into an explorable panorama.
He has been “waiting like months” for an open editor, since Krea 2 and Ideogram were text-to-image only.
Install (ComfyUI):
- Update ComfyUI. Templates → search “Qwen Image 2.1” → three workflows: text-to-image, image edit, remove background. (The workflow files are also downloadable from the model page to drag onto the canvas.)
- Download from
huggingface.co/Comfy-Org/Qwen-Image-2.1:- diffusion model: BF16 (14.2 GB) or INT8 (7.2 GB) →
models/diffusion_models - text encoder (Qwen 3 VL): BF16 (17.5 GB) down to W4A8 (6.3 GB, recommended for low VRAM) →
models/text_encoders - VAE (676 MB) →
models/vae - skip the two prompt-enhancer files (about 9.5 GB each) unless you have VRAM to spare.
- diffusion model: BF16 (14.2 GB) or INT8 (7.2 GB) →
Editing workflow notes: upload up to 10 reference images (bypass or delete unused Load Image nodes; add more with a new Load Image node). The custom size toggle off means the output uses the input image’s size; on means it uses your aspect ratio and megapixels. His car-on-a-forest-road edit took 81 s, which he finds “a bit slower than Flux Klein, which is the previous best edit model.”
Background removal: the remove-background template with a prompt like “remove the background, keep only the woman, and output a PNG image” gave a true transparent PNG in 116 s.
LoRAs: a day after release he found only one, an anime-consistency LoRA for editing (WarmBloodAban/Qwen-Image-2.1-LoRAs). He loaded it at 0.8 strength between the diffusion-model loader and the next node and made a front/back/side character sheet. Stack LoRAs by chaining Load LoRA nodes or using a LoRA-stacker node.
4 GB VRAM: GGUFs. Install the ComfyUI-GGUF extension (city96) from the extensions manager, swap the Load Diffusion Model node for Unet Loader (GGUF), download a GGUF from Abiray/Qwen-Image-2.1-GGUF (smallest 3.19 GB) into models/unet, and select it. His Q3 result was “a similar quality to the full model.” With 6–8 GB, try the INT8 build first, since Q6 and Q8 are not much smaller.
Try It
- If you have ComfyUI, start with MOSS + the 8-step DMD LoRA for drafts and thumbnails: smallest download, fastest loop, good in-image text.
- Use Qwen Image 2.1 for product and catalogue edits: drop a product shot and a scene photo into the edit template and describe the composite; use the remove-background template for transparent PNGs you would otherwise cut out by hand. ^[inferred] Check outputs for detail drift before using them in client work.
- Use MOSS layout boxes for compositions you need exactly (a product on the left, text block on the right, a prop in a corner).
- Match quantization to your card before downloading anything (see the Key Takeaways numbers); on a 4 GB card go straight to the Qwen GGUF route.
- Confirm licence terms on each model card before commercial use. The output-rights point for Qwen comes from a reply the presenter relays, not the licence text.
Implementation
Tool/Service: ComfyUI (free, local), MOSS Image (SFT/Base), Qwen Image 2.1, community LoRAs, GGUFs and fine-tunes.
Setup: ComfyUI Windows portable, updated via update_comfyui.bat; built-in templates for both models; models from Comfy-Org on Hugging Face into models/diffusion_models, models/text_encoders, models/vae, models/loras, models/unet (GGUF).
Cost: free and offline after download. Disk: about 7–17 GB per model file depending on precision.
Integration notes: LoRAs go between the model loader and the KSampler (MOSS) or the next node after the diffusion-model loader (Qwen). CFG 1 disables the negative prompt on both. MOSS edits only by masked inpainting.
Open Questions
- MOSS Image’s maker and licence are not stated in the tutorial. Answered in part 2026-10-30: AI Search’s October 18 news episode (
raw/Quantum_breakthrough_Gemini_4_Argon_MOSS_Image_KLING_4_OpenAI_spy_Muse_Wonder_-_AI_NEWS.md) says “everything is under the Apache 2 license… You can even use this for commercial purposes,” and links the project page atopenmoss.ai/MOSS-Image. Confirm on the model card. - The “4 GB of VRAM or less” claim for MOSS is the presenter’s; he ran it on 16 GB and did not show a low-VRAM run.
- Qwen Image 2.1’s output rights rest on a Qwen reply the presenter paraphrases; the licence text itself says no commercial use without a separate licence.
- No side-by-side against hosted editors (Nano Banana, GPT Image) is shown, only his verdict that Qwen is the first open model to edit in natural language.
Related
- FLUX 3 (Black Forest Labs) — FLUX 3 Image Klein, the previous open edit model he compares speed against
- Nano Banana — the hosted editor both videos use as the reference point
- ChatGPT Images 2.5 Field Test — the hosted alternative for edits and in-image text
- Frontier Model Landscape, October 2026 — Qwen Image 2.1 Turbo, Krea 3 and other October image releases
- Local Video Models in ComfyUI — MiniMax H3 and LTX 2.5 — the same ComfyUI install pattern for video
- Higgsfield — the hosted platform that sponsored the MOSS video