Resource directory
Find what you need for your workflow.
Guides: techniques, setup, and troubleshootingSearch Sweet Tea Pipes, external workflows, models, node packs, and guides in one evidence-linked catalog.
Start with a workflow instead ↗Search results
All resources
20 shown of 41,718
LTX-2.5 image to video with audio
Animate a starting image into one video with generated audio using LTX-2.5. Upload a first frame and describe the scene, action, camera movement, and desired sounds or dialogue. There is no audio-upload input. For a first run, replace the example prompt to match your image and keep 5 seconds, 24 FPS, 16:9, and 0.9 output megapixels. Use an image with a similar aspect ratio when composition matters. Aspect ratio and megapixels determine output dimensions rounded to a 32-pixel multiple. Initial generation uses half those dimensions, followed by 2× spatial latent upsampling and joint video/audio refinement. Increasing resolution, duration, or FPS increases the work required. Enhance prompt selects a rewrite informed by your text and uploaded image. Turn it off to select your original wording. Enhancement length controls the rewrite's text budget, not video duration. Sample enhanced wording uses on/off choices; temperature, probability filters, penalties, and enhancement seed apply when sampling is on. Thinking mode depends on support from the fixed enhancement model, and Use enhancement template selects its built-in template when available. Text-to-video mode bypasses starting-frame injection in both stages. An image upload remains required because preprocessing stays connected. Also turn off Enhance prompt to make generation conditioning come only from your text. Initial and refinement image strengths are independent and inactive in text-to-video mode. The uploaded alpha mask is not connected; transparency is not preserved as an output feature. Enter positive whole-second duration and a whole frame rate from 1 to 120 FPS. Choose duration × FPS divisible by 8, with duration × FPS + 1 at most 1000. At 24 FPS, whole-second durations from 1 through 41 meet these conditions. The default requests 121 frames, so playback can be about one frame longer than five seconds. Generation, refinement, and sampled prompt enhancement have separate seeds. Use -1 to randomize each run. For comparisons with fixed random inputs, set all applicable seeds explicitly and keep enhancement settings unchanged. Both stages share the negative prompt and expose separate video and audio guidance inputs. Keep guidance at 1 for a first run; supported tuning ranges are not established by the supplied evidence. Advanced settings preserve independent samplers and editable comma-separated noise schedules for both stages, plus source-image compression from 0 to 100. Keep their authored defaults initially. There is no separate Steps control: the manual schedules define sampling transitions. Image compression affects the image used for enhancement and conditioning. Export settings offer automatic or explicit container and codec selection, with 8- or 10-bit video assembly. Keep auto and 8-bit for a first run. The source notes that 10-bit supports smoother gradients but may have more limited compatibility. Export settings affect encoding rather than generation. The retained CRF field stays hidden at -1 because its effective quality behavior is not established by the supplied evidence.
Kling text to video
Generate one video from a text description through the Kling partner API. Provider authentication and access are required; generation sends your prompt to the provider and incurs provider charges. No image upload is required. For a first run, use the supplied example prompt or describe your subjects, action, setting, camera movement, and visual style. Keep Kling 3.0 Omni, 16:9 framing, five seconds, and 720p resolution, then run to receive the video result. Audio generation is enabled by default. Include things to avoid in the same prompt when needed. Choose 16:9 for landscape, 9:16 for portrait, or 1:1 for square framing. Duration is in seconds and accepts 3–15; resolution offers 720p, 1080p, and 4k. The model selector also offers Kling video o1. Audio support with o1 is not established by the supplied evidence; turn Generate audio off when trying that model. Model, duration, resolution, and audio settings can affect provider cost. This tool supports one main prompt. Storyboard mode stays disabled because editable segment prompts and durations are unavailable in this interface. Leave Seed at -1 for a randomized rerun seed. A fixed seed can affect rerunning and cache reuse, but provider results remain nondeterministic. The advanced Video output section controls file type, codec, and encoding quality. Leave auto for file type and codec, and -1 for CRF on a first run. These settings control the returned video's encoding; codec-specific CRF behavior is not documented in the supplied evidence.
Luma image transformation
Transform one uploaded source image with a text prompt using Luma Photon-1 and preview the generated result. For a first run, upload your source image, replace the example prompt with your desired subject, appearance, and setting, then run the Pipe. Image weight ranges from 0 to 0.98 and defaults to 0, allowing substantial reinterpretation. Higher values modify the original less; increase the weight when you want the result to stay closer to the source. Leave Seed at -1 to randomize each run. Changing Seed requests a fresh provider run, but reusing a seed does not reproduce an image: Luma results are nondeterministic. The model is fixed to Photon-1 in this Pipe. The image and prompt are sent to Luma through a paid provider call, which requires API access and sufficient credits. Only one source image is supported. Uploaded transparency is not preserved because only the loader's RGB image output is connected.
Image to creative prompt
Create and save a descriptive text-to-image prompt from a reference image. For a first run, upload a Source image, leave the default Request and Response instructions in place, and run. The saved text describes visual details such as subject, style, environment, lighting, composition, colors, and mood. An image is needed while Source image is enabled; the empty upload default requires your own reference. Edit Request to specify what to describe or emphasize. In advanced Response settings, Response instructions independently set writing style, content priorities, and output format. Maximum response tokens limits generated output to a budget of 16–16384 tokens; it is not an exact word count, and a small limit may cut the response short. Context truncation controls oversized input: Auto permits input truncation, while Disabled makes an oversized request fail. For a text-only task, bypass Source image and replace both the image-focused Request and Response instructions with instructions suited to your task. Model is selectable; when using an image, choose a model that accepts image context. This workflow sends the request and enabled image context to OpenAI and uses a provider-priced API call. Its output is text, not a generated image, and it does not guarantee that another image generator will recreate the reference.
Capybara image edit
Edit an uploaded image with written instructions using Capybara v0.1 and receive still-image results. For a first run, upload a source image, replace the example edit instruction with the changes you want, and state which subjects or details should remain. Keep the default sampling settings and generate one image. The prepared source supplies both image conditioning and visual reference features; these do not guarantee that requested details remain unchanged. Width and Height set both the prepared source size and the generated image dimensions in pixels. The default is 1024 × 1024 with a center crop. Center preserves proportions but can remove content near the edges; Disabled stretches the source to fit. Use dimensions divisible by 16, such as 1024 × 1024 or 1360 × 768. Setting one dimension to zero derives it from the other using the source aspect ratio; do not set both to zero, and ensure the derived dimension is also divisible by 16. Resampling method controls source interpolation: Area is suited to reduction, Lanczos to enlargement, and Nearest-exact to pixel-art edges. Images per run requests multiple edited variants. Seed -1 selects fresh randomization each run; a fixed seed reuses the sampling noise. Negative prompt lists unwanted features, and Prompt guidance (CFG) scales positive guidance relative to negative conditioning. Steps sets the sampling step count. Start with the default Euler sampler, Simple scheduler, 20 steps, CFG 6, and sampling shift 7; the advanced controls change sampling behavior. Visual reference crop optionally adds a center crop when encoding reference features, after the source has already been resized. It does not change output dimensions or the separate source-image conditioning branch. Noise schedule fraction controls how much of the sampling noise schedule is used; it is not a direct source-preservation strength control. The output is still images despite the HunyuanVideo conditioning node names. Frame count comes from the prepared image batch rather than an editable duration. There is no mask-based regional edit input, and uploaded transparency is not preserved. Results are returned through the image preview output.
Image to 3D turntable
Create a 360-degree turntable video of a TripoSplat reconstruction from one uploaded image. Adjust output dimensions, orbit frames, background color, and render style: color, clay, depth, or normals. The default 75 frames at 25 FPS produce a three-second video. Negative frame counts reverse the orbit; -1, 0, or 1 render a single view. Disable automatic background removal when using the image's alpha mask. Camera controls adjust framing and projection; bypass the camera stage for automatic framing. Advanced controls tune foreground preprocessing, reconstruction sampling, splat density, intermediate previews, and video encoding. Gaussian counts above 262144 oversample existing points rather than adding detail. Generation seed and point sampling seed control separate random processes.
Seedream image generation
Generate an image from a text description with Seedream and preview the result in Sweet Tea. This Pipe uses the ByteDance provider API and requires configured provider access; generation uses provider billing. It has no reference-image input or separate negative prompt. For a first run, keep the supplied cyberpunk poster prompt or replace it with your subject, composition, visual style and desired text, then run the Pipe. The default model is Seedream 5.0 Pro, and the default output size is 1024 × 1024 pixels. Choose Model and Output size together: recommended size presets depend on the selected model. Select Custom to use Custom width and Custom height; otherwise those fields have no effect. Their stored defaults are 2048 × 2048 pixels. The exposed custom limits are 1024–3136 pixels wide and 1024–2496 pixels high, in steps of two pixels, as supplied for the default Pro model. Prompt reasoning enables the model's prompt-optimization step and is on by default. The source notes that it can substantially increase generation time, especially with 5.0 Pro; this text-only Pipe allows it to be disabled. Seed -1 selects a fresh random seed per run; use a nonnegative seed to control generation randomization. Enable Watermark to add an “AI generated” mark. The selected output returns an image preview through Sweet Tea rather than using the source workflow's permanent image-saving stage.
DALL·E 3 image generator
Generate an image from a text prompt using DALL·E 3 and view the result in the Pipe output. Enter a prompt describing the subject, composition, lighting, and visual treatment, or use the supplied example for a first run with the default square size, vivid style, and HD quality. Choose dimensions in pixels: 1024x1024 for square images, 1024x1792 for portrait, or 1792x1024 for landscape. Vivid favors dramatic, hyper-real imagery; natural favors a less exaggerated appearance. Use the prompt to specify the artistic medium. Quality selects the provider's standard or HD mode independently of output size. This workflow accepts text only; it has no image input or separate negative prompt. It requires authenticated access to the ComfyUI partner API service and sufficient credits, and generation sends the prompt to the external provider. The source marks backend seed support as unimplemented, so a fixed seed does not guarantee repeatable images. The selected output stage returns the generated image for preview rather than saving it through the original ComfyUI SaveImage node.
Brightness and contrast
Create a brighter, higher-contrast image from an uploaded source image. For a first run, upload your image and run with the saved settings: Brightness 13 and Contrast 51. These settings apply an adjustment rather than a neutral preview. To start with no tonal adjustment, set both controls to 0, then change either control and run again. Both controls use a unitless 0 to 100 scale. Brightness adds a linear-light offset from 0 to 1. Contrast uses a multiplier from 1 to 2 around 18% linear middle gray, making values below that pivot darker and values above it brighter. The brightness offset is added after contrast, so the two controls combine in the result. This interface supports brightening and increasing contrast; it does not offer negative adjustments for darkening or reducing contrast. Processing approximates sRGB conversion with gamma 2.2 and clamps adjusted linear RGB values to the available range. Strong settings can clip shadow or highlight detail. Output dimensions follow the source image. Uploaded transparency is not preserved because the image loader's separate alpha mask is not used. This output applies only brightness and contrast; the other adjustments in this family are separate tools.
Vidu Q1 reference to video
Generate one five-second 1080p video from three reference images and a scene prompt using Vidu Q1. References guide the subjects and setting; they are not prescribed start or end frames. For a first run, upload an image into each of the three required reference slots. Replace the example prompt with a description matching your images, explaining how their subjects or elements appear together, what happens, and how the camera moves. Choose landscape 16:9, portrait 9:16, or square 1:1 framing, and leave movement amplitude on Auto initially. Small, Medium, and Large request different amounts of object movement; describe specific actions in the prompt. Leave seed at -1 to randomize each run, or enter a positive integer to request a particular generation seed. The provider also treats 0 as random. Duration is fixed at five seconds, resolution at 1080p, and the model at Vidu Q1. This Pipe uses exactly three connected reference slots even though the provider supports up to seven references. Generation sends your reference images and prompt to the Vidu partner service, requires access to that service, and incurs provider charges. The generated video is returned to Sweet Tea for preview with automatic format and codec selection.
Lip sync video to audio
Create one video with lip movements adjusted to follow an uploaded speech or vocal track. For a first run, upload the source video containing the person to synchronize and upload the audio track their lip movements should follow. Both files are required. Leave the advanced Video output settings at their defaults, then run the Pipe. Video output lets you choose a file type (Auto, MP4, MKV or WebM), codec (Auto, H.264 or AV1) and output encoding CRF. These settings affect the returned video's encoding, not the lip-sync processing. This Pipe uses uploaded audio and offers no recording mode. Lip-sync processing uses the exported node configuration; there are no exposed seed, processing-mode, trimming or duration controls. Behavior when the video and audio lengths differ is not established by the supplied evidence.
Nano Banana image editor
Edit a reference image with written instructions using Nano Banana 2 or Nano Banana 2 Lite, and view the generated image in Sweet Tea. For a first run, upload an image and describe the changes you want, including features that should remain unchanged. The existing example prompt makes an F1 racing poster with a yellow car, replacement lettering and motion effects. Use it with a car image, or rewrite it for your own subject. There is no separate negative prompt; include all constraints in Edit instructions. The default uses Nano Banana 2 Lite at 1K. Select the model before setting resolution: use 1K with Lite, or 1K, 2K or 4K with Nano Banana 2. Auto aspect ratio follows the reference image; another ratio requests different output proportions and composition. Thinking level selects MINIMAL or HIGH reasoning effort. Temperature controls generation randomness from 0 to 2, and nucleus sampling controls the sampling threshold from 0 to 1. Seed -1 requests a new random seed each run. A fixed seed requests best-effort consistency, but identical results are not guaranteed, especially when model or sampling settings change. System instructions provide optional foundational behavior; the supplied default asks the provider to produce an image. This workflow sends the reference image and instructions to the Gemini provider and uses a provider-priced generation call. Generative editing does not guarantee preservation of geometry, lettering or fine details. The uploaded alpha mask is not connected, so transparency preservation is not supported by this graph. Only the generated image is returned; provider text and thinking images are not separate results. The selected output is a Sweet Tea preview, replacing the original file-saving stage.
MoGe depth and normal visualization
Render estimated depth, surface normals or geometry validity from one perspective image. For a first run, upload an image, leave Rendering mode at depth_colored, and run with the existing defaults. The fixed MoGe 2 model estimates geometry, and the Pipe returns one rendered image. Rendering mode selects depth for the alternate depth display, depth_colored for colorized depth, normal_opengl for normals with the +Y-up green-channel convention, normal_directx for normals with the -Y-down convention, or mask for predicted geometry validity displayed as an image. The output is not numeric depth, a mesh or a typed MASK. Validity is predicted from RGB pixels; uploaded transparency is not used, and the upload's alpha channel is not preserved. In Geometry estimation, Detail level ranges from 0 to 9 and defaults to 2. The source identifies 0 as fastest and 9 as most detailed. Horizontal field of view is in degrees, from 0 to 170: leave 0 for automatic recovery from predicted points, or enter a known camera value to set the focal length used for geometry unprojection. Exclude invalid geometry defaults to enabled and sets predicted invalid points and depth to infinity. Disable it to retain raw predicted geometry; the predicted validity mask remains available in mask rendering mode. Automatic resizing defaults to enabled. It checks image width: only images wider than 2048 pixels are resized to a 2048-pixel longer edge before inference. Images at most 2048 pixels wide remain unchanged, including tall images. Disable this option to use the original image dimensions. Resizing can change the rendered output dimensions relative to the upload. Advanced inference options retain the source defaults. Inference batch size sets images per inference call and starts at 4; it does not create additional outputs from this single-image input. Force projection starts enabled. Its precise effect is not documented in the supplied source, so retain the default for a first run. The fixed model file is moge_2_vitl_normal_fp16.safetensors, installed under models/geometry_estimation/. Its retained download source is https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors. Image output transport uses the Sweet Tea nodes package at https://github.com/tea-time-labs/sweet-tea-nodes.
OmniGen2 Image Editor
Create edited images from a source image and natural-language instructions, optionally using a second image as visual reference. For a first run, upload a Source image, replace the crystal-transformation example in Edit instructions with your desired changes, and adjust Negative prompt to describe unwanted features. Leave Additional reference encoding bypassed and keep the supplied guidance and sampling defaults. Image count selects the number of results. Seed -1 randomizes each run; a fixed seed reuses sampling randomness with unchanged settings. Output megapixels sets the approximate pixel budget in millions of pixels. The resized primary image supplies the output dimensions and aspect ratio; independent width and height controls are unavailable. Primary resize method selects interpolation for that preprocessing. Generation starts from an empty latent and uses the source as reference conditioning, so unchanged details are not guaranteed. Uploaded transparency is not preserved. For a second reference, upload an Additional reference image and turn off bypass on Additional reference encoding. Explain in Edit instructions how to use it. The upload is needed only while this encoder is enabled. Reference megapixels and Reference resize method control its preprocessing independently and do not determine output dimensions. Bypassing the encoder restores primary-reference-only editing without requiring the second upload. Text guidance scales the difference between instruction and negative-prompt conditioning with the same enabled references. Reference guidance scales the difference between reference-conditioned negative text and negative text alone; it affects both references together when the second image is enabled. Guidance composition selects regular or nested combination of these terms; the supplied setup uses regular. Advanced sampling controls select the algorithm, noise schedule and step count. Sampling denoise selects the fraction of the schedule used, rather than directly controlling how much of the uploaded image is retained; the supplied setup uses 1. Larger image budgets and batches require more memory. The workflow uses the configured OmniGen2 diffusion model, Qwen 2.5 VL text encoder and ae VAE.
Magnific creative upscale
Enlarge an uploaded image with Magnific's prompt-guided creative upscaler and return the enhanced image. For a first run, upload your source image and replace the example detail prompt with a description of its subject, style and desired details. Keep the 4x scale, standard content preset and supplied tuning defaults, then run the Pipe. Choose 2x, 4x, 8x or 16x enlargement and a content preset suited to your image. Creativity controls reinterpretation; resemblance controls fidelity to the original. Detail definition adjusts definition and detail, while detail intricacy adjusts prompt strength and intricacy per square pixel. These four numeric controls range from -10 to 10. Engine selection offers automatic, magnific_illusio, magnific_sharpy and magnific_sparkle processing choices. Creative upscaling can change image content. Automatic downscaling reduces the input when the requested output would exceed the provider's maximum pixel limit, so final dimensions may be smaller than the selected scale implies. Leave it enabled unless you know the requested output fits that limit. The workflow connects only the loader's IMAGE output and does not preserve uploaded transparency.
Krea 2 Turbo style reference
Generate new images from a scene prompt and one visual style reference using Krea 2 Turbo and its style reference adapter. For a first run, upload a reference image, replace the example Scene prompt with your desired subjects, actions, composition, and medium, and run with the supplied settings: square 1-megapixel output, one image, adapter strength 1, and 8-step Euler/simple sampling. Leave Refine prompt off to use your original text. Put requested visible words in quotes; exact text rendering is not guaranteed. Aspect ratio and Megapixels determine the generated dimensions independently of the reference image. Images per run returns a batch of variants. Larger dimensions and batches require more memory. Image Seed uses -1 for a new random seed each run; a fixed seed controls the initial noise. Enable Refine prompt to expand your text using fixed expansion instructions. Refinement receives text only and does not inspect the reference image. Its advanced token limit and text-sampling controls affect the expanded prompt separately from image generation. Refinement sampling off disables sampled text generation, rather than disabling refinement; use Refine prompt to choose the original text. Thinking mode applies only if the loaded text model supports it. The selected output is generated images, not a separate refined-text result. Adapter strength scales the fixed style reference LoRA. At 0, its contribution is removed while the reference image remains connected to conditioning. It is not an image-editing denoise amount or a guaranteed measure of visual similarity. Generation starts from an empty latent, so the upload is a conditioning reference rather than the starting canvas. Advanced image controls expose steps, sampler, scheduler, prompt guidance, noise schedule fraction, resolution-dependent sampling shifts, and the reference-latent convention. The source uses guidance 1, noise schedule fraction 1, base shift 0.5, maximum shift 1.15, and index_timestep_zero reference conditioning. These controls change sampling or conditioning; the reference-latent options are not visual style presets. Negative conditioning is zeroed internally, with no editable negative prompt. Only the reference loader's RGB image output is connected, so uploaded transparency is not preserved.
FLUX.2 image editing
Generate edited images from an uploaded reference and written instructions using FLUX.2, with standard and Turbo modes. For a first run, upload your source image and replace the example instructions with the changes you want and details to retain. Keep Output megapixels at 1, Image count at 1, Turbo mode off, Standard steps at 20, Guidance at 4, and the Euler sampler. Run the Pipe to receive the generated image batch. Output megapixels resizes the reference while retaining its aspect ratio. The resized reference supplies both reference conditioning and the dimensions used for generation; latent sizing may round dimensions to the model's supported grid. Resize method controls interpolation before reference encoding. Larger images and batches increase resource use. Turbo mode selects the fixed Turbo LoRA and Turbo steps together. Standard steps apply only with Turbo off; Turbo steps and Turbo LoRA strength apply only with Turbo on. The supplied Turbo defaults are 8 steps and LoRA strength 1. Both step controls feed a scheduler supporting 1–4096 iterations. Guidance, sampler, seed, output size, and image count apply in both modes. Seed -1 requests a new random seed each run; use a fixed seed with other settings held constant to compare changes. This generates a new image conditioned on the reference. Unchanged pixels and exact identity are not guaranteed. There is no negative prompt, editing mask, or denoise-strength control. Uploaded transparency is not preserved: only the core image loader's IMAGE output is connected, while its separate MASK output is unused. The supplied dependency evidence marks all four model files and the Sweet Tea output node pack as missing. Install the Sweet Tea node pack from https://github.com/tea-time-labs/sweet-tea-nodes and obtain the fixed model files from their supplied sources: - diffusion_models/flux2_dev_fp8mixed.safetensors: https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/diffusion_models/flux2_dev_fp8mixed.safetensors - text_encoders/mistral_3_small_flux2_bf16.safetensors: https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/text_encoders/mistral_3_small_flux2_bf16.safetensors - loras/Flux_2-Turbo-LoRA_comfyui.safetensors: https://huggingface.co/ByteZSzn/Flux.2-Turbo-ComfyUI/resolve/main/Flux_2-Turbo-LoRA_comfyui.safetensors - vae/full_encoder_small_decoder.safetensors: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors Place these files in the corresponding ComfyUI model directories before running.
Qwen-Image 2512 edge-guided generation
Generate new images from a prompt while following an uploaded reference image's Canny edges. The Pipe returns generated images and an edge-map preview. For a first run, upload a reference and replace the example positive prompt with your desired subject, appearance, lighting and style. Review the negative prompt for unwanted attributes. Keep output size at 1.6 megapixels, control strength at 1 and Lightning mode off, then run. Output dimensions follow the resized reference's aspect ratio. The reference supplies edges and dimensions; generation starts from an empty latent and does not directly preserve source colors or texture. Uploaded transparency is not retained. Lower control strength for more freedom, or set it to zero to remove edge conditioning while retaining the reference-derived dimensions and preview. The Canny thresholds are normalized values: keep the low threshold at or below the high threshold. Lower thresholds admit more edges; higher thresholds focus on stronger boundaries. Inspect the returned edge map when changing them. Control start and end are fractions of denoising progress from 0 to 1; keep start below end. Lightning mode selects the fixed four-step LoRA and separate sampling controls. Its supplied defaults are 4 steps and CFG 1; standard mode uses 50 steps and CFG 4. Only the selected mode's steps and CFG apply. Use positive step counts; the downstream sampler supports 1–10000 steps and CFG 0–100. Euler and the simple scheduler are the supplied choices for both modes. Increase Image count for a batch. Seed -1 randomizes each run; use a fixed seed to compare changes with otherwise matching settings. Advanced controls also expose reference interpolation, denoising amount, sampling shift and Lightning LoRA strength. Denoise defaults to 1; lowering it does not turn this into pixel-preserving image editing. Sampling shift affects both modes. LoRA strength affects Lightning mode only and does not automatically change its steps or CFG. Larger output sizes and batches increase memory use. This interface supports Canny guidance only. Pose or depth guidance would require graph changes.
Flux.2 Klein base 4B image edit
Generate edited images from a reference image and a written instruction using Flux.2 Klein base 4B. For a first run, upload a source image, describe your desired changes and any features to retain, and keep the supplied defaults: 1 megapixel, one image, CFG guidance 5, 20 sampling steps and Euler. The example instruction requests anime styling while retaining identity. Leave the negative prompt empty unless you want to describe attributes to discourage. Seed -1 selects random sampling noise. Output megapixels resizes the reference while following its aspect ratio. The resized image's dimensions also set the generation size. Resize method controls interpolation of that reference before encoding; it does not add a separate upscaling stage. Advanced controls provide image count, negative prompting, CFG guidance, seed, sampling steps and sampler choice. Use a fixed nonnegative seed to reuse the noise setting; changing the reference, instruction or sampling settings can still change the result. CFG combines positive and negative conditioning, with the image reference attached to both branches. The source image supplies reference conditioning, while generation starts from an empty latent. There is no image-to-image denoise-strength control, text-only mode or mask-based editing mode in this tool. Uploaded alpha transparency is not preserved by the connected RGB image path. The output contains generated images. The workflow uses flux-2-klein-base-4b-fp8.safetensors, qwen_3_4b.safetensors as its Flux.2 text encoder, and flux2-vae.safetensors. It uses the base 4B model despite the original subgraph's historical distilled title.
Three-scene story video
Gemini develops the brief into three scene prompts. The first scene also guides Nano Banana 2 to create a reference image, which Kling uses to generate the storyboard video. Disable Source image to work from text alone. The default storyboard has three four-second scenes and generates audio. Video and reference-image aspect ratios are separate source settings; keep them aligned for consistent framing. Scene prompts are generated automatically, and the saved script-format instructions maintain the three-scene structure.
