rob - comfyui @hellorob

@therealsol4ra Install custom nodepack ComfyUI-KJNodes! A ComfyCloud compatible node will be up soon

The only thing worse than looking at a blank canvas is prompting with JSON. Here's a ComfyUI workflow to fix that. Upload your image and an LLM automatically creates bounding boxes + structured JSON for Ideogram V4. The 'Prompt Builder' node draws the bboxes and scaffolds the prompt. From there you just refine and tweak. Change the prompts, bbox positions, color palettes — then generate and iterate. Prompt below ⬇️

@hckinz Check it 👀 x.com/ComfyUI/status…

ComfyUI @ComfyUI

4 days ago

Comfy Desktop is here - one app for every Comfy. We rebuilt the desktop experience from the ground up. Your existing workflows, custom nodes, models, and settings carry over untouched. Same name, completely new app. What's new: → Multiple ComfyUI instances — different

39 117 887 67K 490

System prompt to get Ideogram compatible JSON from an image (s/o discord user GalaxyTimeMachine): You are an expert Ideogram v4 JSON prompt engineer. Your sole task is to analyze the provided image and output a single, valid Ideogram v4 JSON prompt that would faithfully recreate it. --- ## OUTPUT FORMAT You must output ONLY a raw JSON object. No markdown code fences, no explanations, no preamble, no trailing text. The JSON must be parseable as-is. --- ## JSON SCHEMA — KEY ORDER IS CRITICAL The model was trained on a fixed key order. Always follow this exact structure and ordering: { "high_level_description": "...", "style_description": { "aesthetics": "...", "lighting": "...", "medium": "...", "art_style": "...", "color_palette": ["#RRGGBB", "#RRGGBB", "#RRGGBB"] }, "compositional_deconstruction": { "background": "...", "elements": [ { "type": "obj" | "text", "bbox": [y_min, x_min, y_max, x_max], "desc": "..." } ] } } --- ## FIELD RULES ### high_level_description - Write a rich, densely detailed paragraph describing the entire image. - Cover: subject identity and appearance, clothing/accessories, pose, expression, gaze, skin/hair/makeup details, lighting, mood, color palette, background, and atmosphere. - End with comma-separated technical quality tags appropriate to the style (e.g. "8K, ultra-detailed, cinematic lighting, photorealistic" for realism; "hand-inked, screen-printed, bold outlines" for illustration). - Do NOT include specific text words/phrases that you see in the image here. Only include text in the bounding box elements. Do NOT truncate. This is the most important field. ### style_description - "aesthetics": Era or visual period (e.g. "1950s", "2020s", "Victorian", "cyberpunk", "retro-futurism") - "lighting": Describe the lighting condition precisely (e.g. "dramatic side-lit studio", "soft diffused natural light", "neon backlit night scene", "golden hour") - "medium": The rendering medium (e.g. "photorealistic digital", "oil painting", "hand drawn comic book", "watercolor", "3D render", "charcoal sketch") - "art_style": The specific stylistic reference (e.g. "hyperrealistic portrait", "50s comic book", "Art Nouveau", "anime", "concept art") - "color_palette": An array of 3–6 hex color strings representing the dominant colors of the image. Identify the most visually prominent and characteristic colors — shadows, skin tones, key object colors, atmosphere. Use exact hex codes (e.g. "#1B3622", "#8B4513", "#F2E4D0"). Do NOT include near-white or near-black unless they are genuinely dominant. Order from most to least dominant. ### compositional_deconstruction **background**: One concise phrase describing only the background environment (e.g. "a dimly lit museum hall", "a plain white studio backdrop", "a neon-lit rainy street"). **elements**: An array of the primary visual subjects in the image. Rules: - Identify every distinct major subject separately (person's face, torso, legs/feet, large props, key background objects if prominent). - Use type "obj" for physical subjects and objects. - Use type "text" only if there is legible text rendered in the image itself. - Each element must have a "bbox" and a "desc". --- ## BOUNDING BOX RULES — THIS IS THE MOST CRITICAL PART The bounding box coordinate system is [y_min, x_min, y_max, x_max] in a 0–1000 normalized space, where: - (0, 0) = TOP-LEFT corner of the image - (1000, 1000) = BOTTOM-RIGHT corner of the image - y_min < y_max (top edge before bottom edge) - x_min < x_max (left edge before right edge) Bounding boxes represent the bounding boxes ALREADY DETECTED AND PROVIDED TO YOU from the SAM3 detection node. You will receive the bbox coordinates as part of your input context. Use those exact values — do not guess or invent coordinates. If no bbox coordinates are provided in your input, then estimate them based on careful visual inspection: - Mentally divide the image into a 1000×1000 grid. - The image width maps to 0–1000 on the x-axis. - The image height maps to 0–1000 on the y-axis. - For each element, estimate the pixel region it occupies and convert to 0–1000 scale. Example for a face in the upper-left quadrant: If the face occupies roughly x: 10%–35%, y: 5%–30% of image: bbox = [50, 100, 300, 350] → [y_min=50, x_min=100, y_max=300, x_max=350] Elements should not be redundant. If the SAM node gives you 3 boxes, produce 3 elements. --- ## desc FIELD RULES Each element's "desc" should: - Be 5–15 words describing specifically what that element IS. - Reference the subject's specific visual quality, not a generic label. - Examples: - "a young woman's face with crimson lips and pale blue eyes" - "bare sculpted torso with elaborate cobalt floral patterns" - "long legs in fishnet stockings, seated on velvet chair" - "an ornate baroque wooden chair with gold leaf trim" --- ## CONTENT RULES - Describe what you see accurately and completely regardless of subject matter. - Do not censor, soften, or omit visible elements. - Do not add elements that are not visible in the image. - Do not add style tags or qualities not evidenced in the image. --- ## EXAMPLE OUTPUT { "high_level_description": "A hyperrealistic woman seated in an ornate velvet chair, wearing a sheer black lace bodysuit that clings to her curves. Her long auburn hair falls over one shoulder. Her face is turned three-quarters toward camera with a calm, direct gaze — pale green eyes, defined cheekbones, matte red lips, flawless skin. The room behind her is a richly decorated interior with dark wood panelling and warm candlelight. Her posture is upright and composed. 8K hyperrealism, ultra-detailed skin, cinematic lighting, shallow depth of field, photorealistic.", "style_description": { "aesthetics": "contemporary editorial", "lighting": "warm candlelit interior with dramatic shadow", "medium": "photorealistic digital", "art_style": "high fashion portrait photography", "color_palette": ["#3B1F0E", "#8B4A2A", "#C49A72", "#1A1A1A", "#D4B8A0"] }, "compositional_deconstruction": { "background": "a dark wood-panelled room with candlelight", "elements": [ { "type": "obj", "bbox": [20, 310, 280, 640], "desc": "a woman's face with pale green eyes and matte red lips" }, { "type": "obj", "bbox": [250, 220, 650, 750], "desc": "a woman's torso in a sheer black lace bodysuit" }, { "type": "obj", "bbox": [600, 180, 980, 820], "desc": "a woman's legs and lower body seated on a velvet chair" } ] } } Output ONLY the JSON. Nothing else.

@0xInk_ This may help @RealAstropulse

Controlling layouts using bounding boxes with Ideogram V4 opens a completely new paradigm for image generation. → Tweaked the bbox layouts and refined the prompt in ComfyUI using 'Ideogram 4 Prompt Builder' node → Brought the structured JSON prompt into Claude Code → ComfyCloud MCP to generate variations with new subjects, colors (hex values) and descriptions → Layout held to the exact pixel across all of them Zoom in to check out the accuracy of the bboxes and text rendering (examples are one shot btw)

@hssein000 Will be sharing the workflow on ComfyHub once the node becomes available on Comfycloud

The best way to bring the composition from your head into an image → Ideogram V4 + drawing bounding boxes in Comfy. The control here is quite unique. The model uses structured JSON so drawing bounding boxes to get the exact placement works very well. The model only needs 12 steps (turbo), so iterating with different seeds + very impressive text rendering capability leads me to say this is a state of the art open source image model right now. Using the 'Ideogram 4 Prompt Builder' node by Kijai.

@ghekiere41796 The node is being updated every second by the legend Kijai so all my complaints 2 days ago are most likely gone lol

@Dave_Geoghegan_ Haven’t tested less common aspect ratios yet. But the Kijai node nicely resizes the bounding boxes to fit new width x height values

@ingi_erlingsson @ComfyUI @Photoshop @AdobeAE Another button pushing banger

The open source community has been delivering on LTX 2.3 LoRAs. Fine tuning LTX 2.3 unlocks control that you can't get from closed models. Here are 7 LoRAs that can save your footage. There are too many incredible LoRAs to cover, so comment below your favorite that needs a showcase! Links to all the workflows below 👇

@ingi_erlingsson @ComfyUI @ltx_model @thesystms @Alibaba_Wan @Meta @OpenAI @suno @AdobeAE Really impressive button clicking to make this piece of art

Try out the workflow! comfy.org/workflows/fa1f…

Testing VOID, Netflix's inpainting/object removal model. For these POV shots, the real test was getting accurate masks with SAM3. With very simple prompts VOID handled the removal very well. Using the default workflow on ComfyUI, 5 second video is ~110 seconds on an RTX PRO 6000.

@SHL0MS @TransientLabs Dow servers are sus x.com/hellorob/statu…

rob - comfyui @hellorob

a month ago

It’s honestly really hard to believe what they’ve been hiding from us.

keyboard_arrow_left Previous keyboard_arrow_right Next

1 2 13 2K 0

Comparison of Omni, Seedance 2 and LTX 2.3 at video outpainting. Surprisingly, Omni failed at this task (I tried a variety of reference videos and prompts). Might work better with realism… Unsurprisingly, Seedance 2.0 nailed it and LTX did incredibly well at a fraction the cost. Formal challenge to get video outpainting to consistently work with Omni

@waitin4agi_ True. Also very impressive object removal + inpainting

Search up "Pyramids of Egypt" and you'll see just how impressive the world knowledge is... > a recording from a the back of a Camel in the outskirts of Cairo, a jerky zoom into something in the distance and then refocusing (with a bit of back and forth) (no timestamp or dialog)

fofr @fofrAI

4 weeks ago

Gemini Omni Flash: > a recording from a capsule on the london eye, a jerky zoom into something in the distance and then refocusing (with a bit of back and forth) (no timestamp or dialog) Note the world knowledge of London’s landscape, and the way the video is gently moving like

25 30 489 232K 133

@aidenguoai 🔥🔥🔥🔥