Finden Sie, was Sie für Ihren Workflow benötigen.

ArbeitsablaufCivitai

Wan2.1 14B i2v (Native|GGUF) Self-Forcing(lightx2v) Single/Dual Sampling — 🟩v3 (SingleSamp)

Go here for Wan2.2 WF ≈➤ 🍥 Wan 2.2 (GGUF) [i2v / FFLF] + [t2v] Workflow

Vidu Q3 image to video

Animate a starting image into one short video, with optional generated dialogue and sound effects. Upload the required starting image and replace the example chase prompt with action and camera instructions suited to your image. For a first run, keep Vidu Q3 Pro, 720p and the supplied six-second duration. This Pipe uses the Vidu provider integration: configured provider authentication is required, the image and generation request are sent to the provider, and generation uses provider billing. The resulting video is returned to Sweet Tea's video preview output. The prompt may be empty and supports up to 2,000 characters. If you describe timed shots, keep their timing within the selected duration; changing Duration does not rewrite the prompt. Duration supports 1–16 seconds. Pro supports 720p, 1080p and 2K; Turbo supports 720p and 1080p. When switching from Pro at 2K to Turbo, also select a supported resolution. Generate audio is enabled by default. Describe desired dialogue and sound effects in the prompt, or disable it for silent video. Seed -1 randomizes each run; use a fixed seed when desired, without a guarantee of identical provider results.

Vidu Q31024x10248GB VRAM

Wan 2.2 first–last frame video

Generate a silent video guided by start and end images and a motion prompt using Wan 2.2 14B. For a first run, upload a Start image and an End image, replace the saved ice-kitten-to-beast prompt with a description of your scene and transition, and keep the generation and encoding defaults. The default result is 640 × 640 pixels with 81 frames at 16 FPS, approximately 5.1 seconds. The saved negative prompt is Chinese text discouraging visual artifacts, blur, overexposure, static imagery, anatomical errors and clutter; edit it as needed. Each endpoint can be disabled independently. Upload an image for every enabled endpoint; disable its node when omitting it. With only one endpoint enabled, only that boundary is image-conditioned. With both disabled, generation uses the text conditioning without endpoint images. The image loaders supply RGB images only; uploaded alpha masks are unused and transparency preservation is not supported. Width and height use increments of 16 pixels. Frame count follows 1, 5, 9 and so on. Duration equals frame count divided by FPS; changing FPS changes playback timing without adding generated frames. Keep batch size 1 for a single clip: this Pipe exposes one video output, not separate batch-member outputs. Advanced controls include negative text, seed, independent guidance for the initial and refinement stages, sampling algorithms, schedules and model sampling shifts. Seed -1 randomizes the initial noise choice; use a fixed nonnegative seed to hold that choice constant. The default high-noise stage executes steps 0–10 and passes its noisy latent to the low-noise stage, which continues from step 10 through the end of the same 20-step schedule. Refinement runs as part of the complete generation process. If editing schedules, keep both total step counts equal, match their schedulers, and keep the initial handoff step equal to the refinement start step within that schedule. The saved samplers are Euler, guidance is 4 for each stage, and sampling shifts are approximately 8. This variant uses no acceleration LoRAs; the family's four-step Lightx2v variant is a separate configured tool. Output controls select container, codec and CRF. Leave these automatic for the first run; CRF -1 selects automatic quality. Video creation supports 8-bit and 10-bit; the source notes that 10-bit can reduce gradient banding but has player and downstream compatibility limitations. No audio is connected. The graph uses the Wan 2.2 high- and low-noise 14B model files, UMT5 text encoder and Wan 2.1 VAE named in the package source.

Wan 2.21024x10248GB VRAM

Krea 2 Turbo1024x10248GB VRAM

Krea 2 Turbo text to image

Generate an image or batch of images from text with Krea 2 Turbo. For a first run, replace the example positive prompt with your subject, composition, lighting and medium. Keep square 1 MP output, one image, prompt refinement enabled, style LoRA disabled and the supplied 8-step Euler sampling settings, then run the Pipe to receive the image batch. The runtime needs krea2_turbo_fp8_scaled.safetensors, qwen3vl_4b_fp8_scaled.safetensors and qwen_image_vae.safetensors, plus the Sweet Tea image output node. Using the default style additionally requires krea2_darkbrush.safetensors. Model selection for the base generator, text encoder and VAE is fixed in this Pipe. Prompt refinement uses the local text model and fixed instructions to expand or polish your description before image conditioning. Disable refinement to use your original wording. Refinement length, thinking mode and built-in template settings apply only when refinement is enabled. With refinement sampling on, temperature, token filters, repetition and presence penalties, and the refinement seed control sampled wording. Turning sampling off disables those sampling controls. Image noise and sampled refinement have separate seeds. Use -1 for a new random seed each run. To compare runs with stable wording and image noise, fix both seeds and retain the other settings, or disable refinement and fix the image seed. Enable style LoRA to apply the selected adapter and append the style trigger to the original or refined prompt. The default adapter is Darkbrush and the default phrase is 'monochrome ink wash style'. Adapter selection and trigger text are independent: update both when choosing another installed compatible LoRA. LoRA strength controls adapter influence only when style LoRA is enabled. Disabling style removes both the adapter and its appended phrase. Aspect ratio and megapixels jointly determine dimensions, rounded to multiples of eight. Image count sets the output batch size; larger dimensions and batches require more memory. Advanced sampling controls change image generation, with Euler, simple scheduling, 8 steps and CFG 1 supplied as defaults. Keep denoise at 1 for ordinary generation from the empty latent. This Pipe has no uploaded image input or editable negative prompt; its negative conditioning is zeroed.

Qwen-Image1024x10248GB VRAM

Qwen Image 2512 text to image

Generate an image or image batch from text with Qwen Image 2512, using standard sampling or the supplied four-step Lightning configuration. Enter a positive prompt describing your subject, composition, lighting and style. No uploaded image is required. Results are returned through Sweet Tea's image preview output. For a first run, replace the fashion example, review the negative prompt, keep the 1328 × 1328 dimensions and image count of 1, and leave Lightning mode off. Standard mode defaults to 50 steps and CFG 4. The supplied Chinese negative prompt discourages low quality, anatomical distortions, oversaturation, waxy skin, confused composition and distorted text. Enable Lightning mode to select the fixed Lightning LoRA together with the separate Lightning step and CFG settings, initially 4 and 1. Only the selected mode's steps and CFG affect sampling. Lightning strength affects the selected model only in Lightning mode. At CFG 1, negative conditioning does not influence sampling; this applies to the default Lightning configuration. Changing modes can change the result's appearance. Width and height are measured in pixels and must be multiples of 16. Examples include 1664 × 928 for landscape, 928 × 1664 for portrait, and 1472 × 1104 for 4:3. Image count controls the returned batch size. Larger dimensions and batches require more memory. Advanced controls retain separate settings for each mode. Sampling steps must be integers from 1 to 10000, and CFG must be between 0 and 100. Sampler, scheduler, sampling shift, seed and denoise are shared by both modes. For the supplied setup, retain Euler, simple, shift 3.1 and denoise 1. Use a fixed seed to compare settings, or -1 for fresh randomness each run. The diffusion model, text encoder, VAE and Lightning LoRA are fixed. This workflow starts from an empty latent and provides no image editing or reference-image input. Lowering denoise changes sampling but does not add those capabilities.

Illustrious1024x10248GB VRAM

Image upscale and refinement

Refine an uploaded image with prompt-guided diffusion, enlarge it with tiled refinement, and sharpen the final image. The preset Illustrious anime checkpoint, micro-detail LoRA and AnimeSharp upscaler target illustrated imagery. Generative refinement can change details, textures and anatomy. For a first run, upload an illustration, keep the default prompts, 2× upscale factor and refinement strengths, then run. Both diffusion stages share the positive and negative prompts. Add a description of your subject to the positive prompt for specific guidance. The default positive prompt includes addmicrodetails. Negative terms discourage traits; they do not guarantee removal. Initial denoise controls changes before enlargement. Tile denoise controls changes during the enlarged redraw. Lower values retain more incoming structure; higher values allow stronger reinterpretation. Each stage has independent seed, steps, CFG, sampler and scheduler controls. Seeds default to -1 for fresh randomization. Set both seeds to fixed nonnegative values when comparing settings with repeatable noise. Upscale factor multiplies image width and height; tile width and height specify processing regions in pixels rather than output dimensions. Tile pattern changes redraw order. Tile mask blur softens redraw boundaries, and padding adds surrounding context. Larger tiles, padding and tile batches require more memory. Tile batch size affects the main redraw, not seam correction. Uniform tiles keeps boundary tiles at the specified dimensions; disabling it allows smaller boundary tiles. Tiled decoding applies to the enlarged refinement stage, not the initial decoder. Seam correction is enabled by default and performs additional generative processing. Its default denoise is 1, so seam regions can change substantially. Lower seam denoise to retain more existing detail, or disable seam correction to skip that processing. Seam band width applies only to Band Pass mode; seam mask blur and padding control blending and surrounding context. The advanced detail LoRA controls independently adjust diffusion-model influence in both passes and positive-prompt encoder influence. Negative conditioning uses the encoder without this LoRA. Set both LoRA strengths to zero to disable its influence. The checkpoint, LoRA file and upscaler model are preset rather than selectable. Final sharpening has amount, radius and sigma controls. Reduce amount if edges develop halos, or bypass sharpening for the unsharpened upscale. The result is one final image; intermediate images are not saved. Uploaded transparency is not preserved.

FLUX Virtual Try-On1024x10248GB VRAM

FLUX virtual try-on

Generate a virtual try-on image of a person wearing clothing from a garment reference. For your first run, upload a clear person photo showing the area to dress and one garment reference image, such as a product shot or flat lay. Keep the default Styling instructions, which request replacement of the full outfit, then run and inspect the returned image preview. Change Styling instructions to describe how the garment should be worn, such as upper body only or tucked in. You can request retention of the person's face, pose or garment details, but inspect the result for changes; these instructions do not guarantee preservation. The garment input accepts one image. For several pieces, supply an already combined reference and identify the pieces in your instructions. This Pipe does not combine separate garment uploads. Seed defaults to -1 for a new random seed on each run. Enter a fixed seed to control generation variation; keep the images and instructions unchanged when comparing seed choices. Generation uses the BFL provider, sends the input images and instructions to that provider, and incurs provider-priced usage. The selected output is an image preview; this workflow does not automatically save an image file.

HiDream-I1 Fast1024x10248GB VRAM

HiDream I1 Fast image generator

Generate an image or batch of images from a text description using the fixed HiDream I1 Fast FP8 model. No uploaded media is required. For a first run, replace the example prompt with your subject, composition, lighting and style. Keep the 1024 × 1024 dimensions and batch size 1, then run to receive the generated image. Width and height are in pixels and must be multiples of 16. Change both to choose an aspect ratio; increase batch size to receive several images together. Larger dimensions and batches require more memory. Seed -1 randomizes each run. Set a specific seed to compare prompt or sampling changes with the same starting noise. Advanced sampling defaults are 16 steps, CFG 1, LCM sampler, normal scheduler, sampling shift 3 and denoising strength 1. The negative prompt has no effect at CFG 1. Raising CFG above 1 applies negative guidance and changes prompt influence. Sampler, scheduler, steps and sampling shift offer alternative sampling configurations; the template defaults provide a starting point. Denoising strength below 1 uses only part of the noise schedule. Because generation starts from an empty latent, it does not preserve a reference image. The diffusion model, supporting text encoders and VAE are fixed. This tool provides text-to-image generation, without image uploads or model selection.

Wan 2.1 VACE 14B1024x10248GB VRAM

Wan 2.1 video transformation

Generate a silent, prompt-guided video conditioned on an uploaded source video using the fixed Wan 2.1 VACE 14B model. First run: upload a source video, replace the example Video prompt with your desired scene, appearance and action, and leave Reference image and Edge guidance bypassed. Keep the remaining defaults initially: 720 × 720 pixels, 81 frames, 16 FPS, latent batch size 1, and the enabled CausVid LoRA with 4 sampling steps and CFG 1. The default output lasts about 5.1 seconds. The Negative prompt contains supplied Chinese exclusions and can be edited in advanced settings. To add appearance guidance, upload a reference image and enable the Reference image stage. To condition on source edges instead of original source frames, enable Edge guidance; its low and high Canny thresholds apply only in that mode. VACE conditioning strength controls the connected visual conditioning separately from prompt guidance. Frame count and output FPS are independent: duration equals frame count divided by FPS. Width and height are measured in pixels. Source audio and source FPS are not transferred, and the result is newly generated video rather than an exact preservation of the input. The CausVid LoRA can be bypassed. When doing so, the supplied guidance recommends changing Sampling to 20 steps and CFG 6; bypassing does not change those controls automatically. Seed -1 randomizes sampling noise. Advanced settings also retain the sampler, noise scheduler, denoising amount, sampling shift, latent batch size and LoRA strengths. Video playback offers 8- or 10-bit assembly. The supplied source notes possible player and downstream compatibility limits for 10-bit. Video output provides container, codec and CRF choices; automatic container and codec selection and CRF -1 retain transport defaults. Keep latent batch size 1 for the documented single-video use, since separate batch-item output files are not exposed. This Pipe returns one video. The model is fixed; alternate model branches, animated WebP output and edge previews are not selectable outputs.

FLUX.1 Krea [dev]1024x10248GB VRAM

Remove backgrounds with BiRefNet

Upload an image to remove its background with BiRefNet and return a transparent image. A segmentation mask is generated internally for alpha compositing and is not a separate output.

Magazine packaging presentation

Create a magazine packaging presentation in two generative stages: first design a flat cover, then reinterpret that cover as a physical magazine in packaging. Choose this tool when you want a presentation image showing materials, closures and a setting around the magazine. For a first run, upload a photo, replace the masthead and cover lines in Cover prompt, and leave the packaging instructions at their defaults. The default result requests a magazine inside a lightly glossy, transparent plastic envelope with a string-and-button closure on a black background. To generate without a photo, bypass Source image and describe the cover imagery as well as its typography in Cover prompt. Cover model changes only the cover-generation stage; Packaging model changes only the presentation-generation stage. Each independently supports gemini-3-pro-image-preview or Nano Banana 2 (Gemini 3.1 Flash Image), with the source Pro selection retained by default. Cover aspect ratio and resolution shape the intermediate cover. Packaging aspect ratio and resolution shape the returned presentation; auto follows the generated cover's aspect ratio. The 2K and 4K settings use the provider's native upscaler. Every run includes upstream cover generation before packaging generation; uploading an existing cover does not skip that first operation. Packaging prompt controls the material, closure, reflections, background and composition. Both stages generate images rather than applying a precise layout or compositing operation, so review spelling, cover fidelity and physical details. Transparent plastic describes the depicted material, not an output alpha channel. Seeds are independent: -1 randomizes each stage on each run, and fixed values request best-effort repeatability without guaranteeing identical results.

FLUX.1 Krea image generation

Create images from text using the fixed FLUX.1 Krea dev model, CLIP-L and T5 text encoders, and VAE. No uploaded image is required. For a first run, describe your subject, composition, lighting, and style in the positive prompt. Keep the default 1024 × 1024 size, one image, 20 steps, Euler sampler, and Simple scheduler. The supplied portrait prompt is an example; replace it for another subject. Width and height are measured in pixels and must be multiples of 16. Image count generates a batch with the same prompt and dimensions. Larger dimensions and image counts require more memory and computation. Open Sampling for advanced controls. Seed defaults to -1, which randomizes each run. Use a fixed seed when comparing prompt or sampling changes with other settings unchanged. Steps controls the number of denoising steps; sampler selects the sampling algorithm, and scheduler distributes noise levels across those steps. CFG defaults to 1; higher values amplify the difference between positive conditioning and zeroed negative conditioning. This graph has no editable negative prompt or separate FLUX guidance control. Denoise controls the fraction of the sampling schedule used, from 0 to 1. Keep its default of 1 for full generation. The workflow always starts from an empty latent, so reducing denoise does not enable image editing or preserve an existing image. Model and encoder selection are fixed in this tool.

Z-Image Turbo1024x10248GB VRAM

Z-Image upscale and refine

Upload an image to produce one resized, RealESRGAN-upscaled and Z-Image Turbo diffusion-refined image. The selected result includes the entire processing chain; there is no separate upscale-only output or refinement bypass. For a first run, upload your source image and describe its subject, composition and desired details in the positive prompt. Keep the source defaults: 1 megapixel working size, 0.5 final scaling, 0.33 refinement strength, five sampling steps and prompt guidance (CFG) 1. The default prompt, “masterpiece, 8k,” is a minimal starting point; its text does not determine resolution. Working size sets the pixel budget before the fixed 4× RealESRGAN upscaler. Final scaling then multiplies the upscaled width and height before diffusion refinement. Approximate output megapixels equal working megapixels × (4 × final scaling)². The defaults therefore produce approximately 4 megapixels. Aspect ratio follows the upload, subject to image and latent rounding. This is not a fixed enlargement relative to the original upload; the working-size stage can also shrink a large source. Refinement strength controls how much diffusion can reinterpret the resized image. Lower values maintain more source structure; higher values allow stronger changes. Use a prompt that describes the image and the changes you want. Refinement can alter details rather than merely recover them. The negative prompt has no effect at CFG 1. It becomes useful for describing unwanted attributes when experimenting with guidance above 1. Steps, sampler and scheduler change refinement behavior. Advanced model sampling shift is a separate noise-schedule control, retained at the source default of 3 for the first run. The model setup is fixed to Z-Image Turbo with the Qwen text encoder, ae VAE and RealESRGAN_x4plus upscaler. Seed -1 randomizes each run. Use a fixed seed to compare settings with the same noise initialization. The Pipe uses the loader's RGB image output and does not preserve uploaded transparency.

Create a HeyGen talking-head video

Generate one talking-head video with a HeyGen avatar speaking your written script. For a first run, replace the demo script, keep Annie Lounge Standing Side and (avatar's default voice), and leave resolution at 720p, aspect ratio at auto, and speech speed at 1. Avatar and voice are text fields for supported HeyGen selections; this Pipe does not supply selection lists. Custom avatar and voice ID fields accept existing HeyGen identifiers. Leave them empty when using the regular selections. These fields do not create or train an avatar or voice. Choose 720p or 1080p resolution and an automatic, landscape, portrait, square, 4:5, or 5:4 aspect ratio. Background color accepts a HeyGen-supported color value; leave it empty for the provider's default behavior. Speech speed is a multiplier from 0.5 to 1.5: 1 is normal speed, lower values slow speech, and higher values accelerate it. The advanced Video output section offers automatic, MP4, MKV, or WebM file type and automatic, H.264, or AV1 codec selection. Leave both on auto for a first run. Seed -1 randomizes each run; a fixed seed does not guarantee identical provider-generated video. This retained workflow uses the automatic avatar engine and script speech only. It has no audio upload, reference-image input, separate language selector, or selectable alternative engine or speech branch.

1024x10248GB VRAM