How Digital Creators Use Dezgo's Controlled Text-to-Image Mode to Preserve Character Poses and Structural Outlines

In pure text-to-image generative workflows, visual artists and concept designers frequently encounter a frustrating bottleneck: prompt drift. Describing complex physical postures, anatomical foreshortening, or tight architectural compositions solely through text often yields mutated limbs, shifted camera angles, and broken character silhouettes.

Dezgo resolves this fundamental limitation through its dedicated Controlled Text-to-Image mode (accessible on the web app and programmatically via the POST /controlnet API endpoint). Powered by ControlNet neural adapters built on top of the Stable Diffusion 1.5 ecosystem, this feature enables creators to use reference imagery as structural scaffolding. By pairing conditioning models such as OpenPose, Canny Edges, and Depth with granular control weights, artists can lock down poses and linework while allowing text prompts to completely restyle characters, lighting, textures, and environments.

The ControlNet Mechanism on Dezgo

Standard diffusion models convert text tokens into spatial cross-attentions, attempting to guess both the semantics (what something is) and the geometry (where everything sits) simultaneously. ControlNet splits this burden by introducing an auxiliary neural network branch that reads explicit structural feature maps from an initial reference image (init_image) and injects them directly into the diffusion process.

Dezgo integrates several specialized ControlNet models under its control_model parameter:

1. OpenPose (openpose_full): Detects human skeletal joints, limb angles, head orientation, and hand positions from a reference image. It isolates kinetic posture completely from clothing, identity, hair, and backgrounds, making it the premier choice for character turnarounds, action scenes, and expressive poses.
2. Canny Edges (canny): Employs high-frequency edge detection algorithms to trace sharp silhouette borders and structural outlines. It binds the AI strictly to contour lines, making it indispensable for vector-to-render workflows, product design concepts, and hard-surface machinery.
3. Depth (depth): Constructs a monocular z-depth gradient map where white pixels represent foreground elements and black pixels represent distant backgrounds. Instead of constraining specific line contours, Depth preserves three-dimensional spatial volume, camera perspective, and room staging.
4. Additional Structural Models: Dezgo also supports lineart (tuned for manga inks and line illustrations), scribble (for turning loose hand-drawn doodles into full paintings), mlsd (specialized in straight architectural lines), and normalbae (normal map surface orientation).

ControlNet Models at a Glance

Step-by-Step Workflow: Controlled Generation in Dezgo

Digital creators can deploy Dezgo's Controlled Text-to-Image tools directly inside a browser without hardware setup, or through API requests. Below is the production workflow for locking down poses and outlines.

1. Access the Dedicated Tool

Navigate to Dezgo's navigation menu under Image Tools and select Controlled Text-to-image (dezgo.com/app/controlnet). This interface exposes structural conditioning settings that are absent in the standard text-to-image generator.

2. Supply the Reference Image (Init Image)

Upload a reference JPEG or PNG to the Init image drag-and-drop zone.

• Resolution Considerations: Dezgo's underlying ControlNet pipeline operates on the Stable Diffusion 1.5 architecture, which natively targets 512×512 pixel latent spaces. Uploaded images are automatically resized.
• Image Framing: Ensure your reference image aspect ratio roughly aligns with your chosen output aspect ratio (Portrait, Square 512×512, or Landscape) to avoid unwanted subject cropping or compression.

3. Select the Target Control Model

Choose the structural extraction model based on your creative objective:

• Use OpenPose when you need a specific human stance (such as a martial arts jump or a subtle seated posture) but want to change clothing, gender, or species entirely.
• Use Canny Edges when transferring an existing black-and-white ink sketch or preserving the silhouette of an intricate prop or helmet.
• Use Depth when establishing an atmospheric camera angle across foreground and background assets.

4. Configure the "Control Preprocess" Switch

In the More options collapsible drawer, locate Control preprocess (enabled by default):

• Enabled (Yes / true): Dezgo's server automatically extracts the pose skeleton, Canny outline, or depth gradient from your uploaded RGB photo or sketch.
• Disabled (No / false): Use this when uploading a file that is already an extracted ControlNet map—for example, a pre-rendered OpenPose color skeleton from Blender or Posemaniacs, or an inverted black-and-white depth pass from 3D modeling software.

5. Calibrate the Control Scale Slider

The Control scale slider (spanning 0% to 100%, or 0.0 to 1.0 in API requests) dictates the degree of structural authority ControlNet holds over the diffusion process:

• 100% (1.0): Rigid lock. The final image strictly mirrors the reference geometry. If using Canny or Scribble, the AI will fight to justify every line, sometimes generating rigid artifacts or unnatural anatomy if the prompt contradicts the reference.
• 60% – 80% (0.60 – 0.80) [Recommended Sweet Spot]: Balanced flexibility. The character retains the intended limb angles and composition, while the diffusion model smoothly handles fabric folds, hair physics, and dynamic lighting without visual stiffness.
• Below 40% (0.40): Loose inspiration. Useful when you only want a suggestive compositional gesture and prefer the prompt to dictate anatomy.

6. Select the Base Diffusion Checkpoint

Dezgo allows users to choose from dozens of customized SD 1.5 checkpoints via the Model selector:

• Photorealism & Cinematic Portraits: absolute_reality_1_8_1, cyberrealistic_3_3, or realistic_vision_5_1.
• Anime, Manga, and Stylized Characters: anything_5_0, waifudiffusion_1_4, or ghostmix_v2.
• Fantasy Concept Art & Painterly Illustration: dreamshaper_8, deliberate_2, or realdream_12.

7. Balance Prompts, Guidance, and Sampling

• Prompt Writing: Focus your positive prompt on style, lighting, material rendering, and subject attributes (e.g., "cyberpunk warrior in illuminated neon armor, volumetric mist, rim lighting, 8k concept art"). Do not over-explain limb coordinates—ControlNet handles spatial layout.
• Negative Prompting: Maintain Dezgo's default quality negative prompt tokens (ugly, poorly drawn hands, poorly drawn feet, extra limbs, bad anatomy, deformed) to prevent diffusion defects in complex joint areas.
• Guidance (CFG Scale): Keep the slider between 6.0 and 8.5 (default is 7.0). Pushing guidance past 10.0 often clashes with ControlNet conditioning, resulting in burnt edges or high-contrast chromatic fringing.
• Sampler: The default dpmpp_2m_karras provides crisp rendering at low-to-moderate step counts (14–25 steps).

Practical Production Scenarios

Scenario A: Character Consistency Across Multiple Poses

When building an animation model sheet or comic narrative:

1. Generate an initial hero concept in standard Text-to-Image to establish character identity.
2. Note down the prompt, negative prompt, and seed.
3. Switch to Controlled Text-to-Image, set the model to openpose_full, and upload new posture poses (e.g., running, crouching, swinging a sword).
4. Lock the same base model and append relevant character LoRAs (Dezgo supports stacking up to 4 CivitAI LoRAs simultaneously).
5. The output preserves the exact physical action while maintaining the character's costume, armor, and aesthetic palette.

Scenario B: From Rough Sketch to Finished Concept Art

For illustrators transitioning pencil sketches into production art:

1. Upload a scanned digital line drawing into the Init Image slot.
2. Select canny or scribble.
3. Set Control Scale to 0.70.
4. Prompt for environmental lighting and digital painting mediums (e.g., "matte painting, atmospheric daylight, lush fantasy forest background, highly detailed oil texture").
5. The generated concept strictly retains the proportions of your sketch while photorealistically rendering ambient occlusion, textures, and cast shadows.

Commercial Access and API Integration

Dezgo operates on a friction-free tiering structure:

• Free Web Tier: Anyone can test Controlled Text-to-Image without an account or credit card. Generation parameters are constrained to standard step counts (around 14 steps) and single image outputs.
• Power Mode / Prepaid Balance: Creators who log in can activate Power Mode by pre-funding a pay-as-you-go balance (no recurring monthly commitment required). This unlocks parallel multi-image batching (Count), elevated sampling steps (up to 150), Real-ESRGAN 2x upscaling, and seed preservation.
• Dezgo Developer API: Studios can integrate the POST /controlnet endpoint directly into game engines, Discord bots, or creative pipelines:
  • Authentication requires the X-Dezgo-Key header.
  • Inputs accept JSON payloads (with base64 image strings) or multipart/form-data.
  • Compute cost scales transparently: due to dual-network evaluation (the base diffusion U-Net plus the ControlNet feature extractor), /controlnet requests cost roughly twice the rate of basic /text2image requests (e.g., $0.0038 per 512×512 inference at baseline steps).

Summary Checklist for Best Results

• Match Aspect Ratios: Align the orientation of your reference image with your canvas dimensions to avoid warping.
• Dial in the 0.70 Sweet Spot: Start at a 70% control scale. Only increase to 100% if minor posture drift occurs, and lower to 60% if the image appears stiff or artificial.
• Clean Hand References: OpenPose captures fingers best when hands are clearly silhouetted against backgrounds rather than clenched or obscured.
• Leverage SD 1.5 Ecosystem: ControlNet on Dezgo is optimized around Stable Diffusion 1.5 checkpoints; take advantage of its rich library of specialized portrait and anime models.

References

• Dezgo - Free Online AI Image/Video Generator, Apps & Tools
• Controlled Text-to-image - Free Online AI Image Generator - Dezgo
• AI Image Generator API - Flux/Stable Diffusion - Dezgo
• API Reference - AI Image Generator API - Flux/Stable Diffusion - Dezgo
• Overview - AI Image Generator API - Flux/Stable Diffusion - Dezgo